Ensuring data confidentiality via plausibly deniable encryption and secure deletion – a survey

Ensuring data confidentiality via plausibly deniable encryption and secure deletion – a survey Ensuring confidentiality of sensitive data is of paramount importance, since data leakage may not only endanger data owners’ privacy, but also ruin reputation of businesses as well as violate various regulations like HIPPA and Sarbanes-Oxley Act. To provide confidentiality guarantee, the data should be protected when they are preserved in the personal computing devices (i.e., confidentiality during their lifetime); and also, they should be rendered irrecoverable after they are removed from the devices (i.e., confidentiality after their lifetime). Encryption and secure deletion are used to ensure data confidentiality during and after their lifetime, respectively. This work aims to perform a thorough literature review on the techniques being used to protect confidentiality of the data in personal computing devices, including both encryption and secure deletion. Especially for encryption, we mainly focus on the novel plausibly deniable encryption (PDE), which can ensure data confidentiality against both a coercive (i.e., the attacker can coerce the data owner for the decryption key) and a non-coercive attacker. Keywords: Data confidentiality, Plausibly deniable encryption, Secure deletion Introduction Spahr LLP: State and local governments move swiftly to Modern computing devices (e.g., desktops, laptops, smart sue equifax 2017); Third, it will directly violate regulations phones, tablets, wearable devices) are increasingly used like HIPAA (Congress 1996), Gramm-Leach-Bliley Act to process sensitive or even mission critical data. Protect- (Congress 1999), and Sarbanes-Oxley Act (Sarbanes and ing confidentiality of those sensitive data is of paramount Oxley 2002). The data confidentiality should be ensured importance because: First, data leakage will endanger data not only during their lifetime (i.e., the data are preserved owners’ privacy. For example, the data leakage of iCloud in the devices), but also after their lifetime (i.e., the data in 2014 disclosed almost 500 private pictures of various have been removed from the devices). This is because, celebrities (Cbsnews: Apple’s celebrity icloud leak proba- by recovering sensitive data which have been deleted, the bly has mundane causes 2014). Second,itwillruinrepu- attacker can achieve a similar gain compared to success- tation of businesses. For example, Equifax data breach in fully attacking the confidentiality of the data being pre- July 2017 caused a leak of 145,500,000 consumer records; a served in the devices. For example, by recovering a naked few local governments like cities of Chicago and San Fran- picture deleted by a victim, the adversary can still use it to cisco, as well as the Commonwealth of Massachusetts, embarrass the victim or ask the victim for ransom money. have filed enforcement actions against Equifax (Ballard Correspondingly, the research efforts for protecting data confidentiality can be divided into two categories: encryption and secure deletion. Encryption can protect *Correspondence: jiashijie@iie.ac.cn; bchen@mtu.edu Data Assurance and Communication Security Research Center, Chinese confidentiality of the data stored at rest by transform- Academy of Sciences, Beijing, China ing them into another format using some secrets (e.g., State Key Laboratory of Information Security, Institute of Information keys), such that the adversary is not able to correlate Engineering, Chinese Academy of Sciences, Beijing, China the transformed format to the original format without Department of Computer Science, Michigan Technological University, Houghton, USA obtaining the secrets. All types of existing encryption Full list of author information is available at the end of the article mechanisms like symmetric encryption and asymmetric © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Zhang et al. Cybersecurity (2018) 1:1 Page 2 of 20 encryption can achieve the aforementioned security prop- against both the coercive adversaries and the non-coercive erty. Secure deletion is to ensure that once the sensitive adversaries. We summarize the research for PDE the- data are deleted, the probability of recovering them is ory and systems (including both the desktop systems and negligibly small. This requires special techniques to com- the mobile systems). For the second case, we summa- pletely destroy data, eliminating any traces which may lead rize secure deletion approaches in various storage media to a full/partial data recovery. includingharddiskdrives(HDDs)and NAND flashmem- To protect confidentiality of the data stored in a com- ory. We also outline the new direction of secure deletion puting device, conventional encryption may not work approaches by eliminating impacts of operation history. when both the computing device and the device’s owner are captured by an attacker, since the attacker can coerce Organization. In “Background”section,weintroduce the owner to disclose the secret (i.e., a coercive adver- the background knowledge of flash memory, the architec- sary). Once the secret is disclosed, the transformed format ture of a storage system, PDE as well as secure deletion. created by encryption will be reversed, and the sensitive In “Models and assumptions” section, we unify the adver- data will be leaked. A novel encryption technique, plau- sarial model for both PDE and secure deletion, and also sibly deniable encryption (PDE) (Canetti et al. 1997), has summarize the assumptions required by PDE. We then been designed to complement the traditional encryption summarize the literature for PDE in “Protecting data con- to handle such a coercive attack. The high-level idea of fidentiality against coercive adversaries via PDE”section PDE is, the original sensitive message is encrypted into and for secure deletion in “Ensuring confidentiality of ciphertexts in a special way, such that during decryption, the deleted data via secure deletion”section,respectively. if a true key is used, the original sensitive message can In “Future directions”section,wediscuss afew future be recovered, but if a decoy key is used, a plausible mes- directions of PDE and secure deletion. We conclude in sage will be generated. Therefore, upon being coerced, “Conclusion”section. the owner can simply disclose the decoy key to protect confidentiality of the original sensitive message. Background To protect confidentiality of the data deleted from Flash memory a computing device, the deleted data should be made Flash memory is a solid-state non-volatile computer stor- completely unrecoverable. Conventionally, this is ensured age medium that can be electrically erased and repro- by carefully over-writing the storage medium storing grammed. It can eliminate mechanical limitations and the data using garbage information (Joukov and Zadok latency of hard drives, achieving a much higher I/O 2005; Wei et al. 2011; Garfinkel and Shelat 2003;Sun throughput with a much lower power consumption. et al. 2008; Gutmann 1996) or deploying encryption Therefore, it gains popularity in main-stream mobile using ephemeral keys (Perlman 2005a, b,Geambasu devices like smart phones, tablets, wearable devices. In et al. 2009,Tangetal. 2012, Reardon et al. 2012,Zarras addition, a lot of high-end laptops like Apple MacBook use et al. 2016). This unfortunately was shown to be insuf- flash memory as external storage. Even cloud providers ficient, since past existence of the deleted data will cre- allow their users to choose solid state drives (SSDs) as ate impacts on both the data organization (Bajaj and the underlying storage media (Amazon: New SSD-Backed Sion 2013b) and the other data which have not been Elastic Block Storage 2016). The flash memory being used deleted (Bajaj and Sion 2013a). Those impacts can then as the storage medium is mainly NAND flash. NAND be utilized by the adversary as an oracle to derive sen- flash stores data using an array of memory cells, which are sitive information about the deleted data. In the worst grouped into pages (each page can store 512-byte, 2KB, case, the adversary is able to completely recover the data or 4KB data), and multiple pages are further grouped into being deleted (Chen et al. 2016). Therefore, recent secure blocks (can contain 32, 64, or 128 pages). deletion approaches focus on eliminating those impacts Compared to traditional mechanical disks, NAND flash (Bajaj and Sion 2013a, b,ChenandSion 2015;Chenand has several unique characteristics. First, NAND flash has Sion 2016;Jia et al. 2016). an erase-before-write design. Specifically, to overwrite In this work, we aim to conduct a thorough literature a flash page, the page needs to be erased before any review on data confidentiality protection. We believe that new data can be programmed to it. Second, the unit of data confidentiality should be always ensured no mat- read/program of NAND flash is a page, but the unit of terthedata arepreserved in thecomputing device or erase is a block, which consists of multiple pages. The have been removed. Therefore, our survey covers tech- first and the second special characteristics of NAND flash niques being used to protect data confidentiality for both make it expensive to perform an in-place update. There- cases: 1) data are stored in the devices; 2) data have been fore, flash memory usually prefers an out-of-place update. removed from the devices. For the first case, we mainly Third, each flash block has a finite number of program- focus on PDE which can provide confidentiality guarantee erase (P/E) cycles. In other words, a flash block will be Zhang et al. Cybersecurity (2018) 1:1 Page 3 of 20 worn out if the number of programs/erasures performed over it exceeds a certain threshold. Last, NAND flash is vulnerable to read/program disturb (Cai et al. 2015). In other words, frequently reading/writing the same flash location may corrupt the data stored nearby. Due to read/program disturb (Cai et al. 2015), NAND flash usu- ally requires computing an error-correcting code (ECC) for each page, and storing the parity in the “spare area” of the corresponding page. In general, the raw flash can be managed through either flash translation layer (FTL) or flash-specific file system. – Flash translation layer. To be compatible with traditional block-based file systems (e.g., EXT4), flash memory can be emulated as a block device by exposing a block-based access interface, which is the most popular form of flash-based products (e.g., SSDs, eMMCs, USB sticks). This is usually achieved Fig. 1 The architecture of storage systems by introducing special firmware, flash translation layer (FTL), between the file system and the raw flash. FTL can translate the logical block addresses to the underlying physical flash addresses, providing a block The lowest layer is the physical medium layer, where access interface to upper layers. data are actually stored, e.g., HDDs or NAND flash. The – Flash file systems. Another alternative of using raw physical storage medium is always accessed through a flash is to directly build a flash-specific file system controller. The basic function of the controller is to trans- over it. A flash file system is a file system optimized late data format on the physical storage medium (e.g., specifically for flash memory. Popular flash file electrical voltage) into another format (e.g., binary values) systems include YAFFS (Robust Flash Storage: understandable by upper layers. The controller offers a YAFFS 2002), UBIFS (Memory Technology Devices: standardized and well-defined hardware interface, e.g., UBIFS 2015), JFFS2 (Sourceware: Jffs2 2003), and ATA (Team work systems: Advanced technology attach- F2FS (Lee et al. 2015). Note that flash file systems ment 2017) and SCSI (Incits: Scsi storage interfaces 2016), become less popular nowadays. Most of the recent which allows data to be read from/written to the physical mobile devices are only designed to be compatible storage medium. The HDD adopts in-place updates, and with FTL-based flash devices, and usually do not hence its controller usually consistently maps a logical allow directly accessing the raw flash. For example, block address to a certain storage location on the physical the Google Nexus 6P Android phone uses eMMC storage medium. On the contrary, NAND flash prefers cards as storage media, and only the old Android out-of-place updates due to its special features, and is usu- phones like Nexus One and Nexus S allow directly ally managed through FTL or a flash specific file system accessing the raw flash. (“Flash memory” section). Device drivers are to consol- idate access to different types of hardware by exposing Special internal management (being incorporating into a common simple interface in the form of software. The FTL or flash file systems) is usually required to han- block device driver interface allows reading and writing dle the characteristics of flash memory, which may easily of blocks in logical addresses. The block device driver can lead to deniability compromise or data leakage, making beused on topofanHDD controlorNANDflash being it challenging to provide deniability/secure deletion on encapsulated by FTL. The memory technology device flash-based storage systems. (MTD), another type of device driver, is used to access raw NAND flash memory directly. MTD permits reading The architecture of a storage system and writing, but blocks must be erased before being The storage media like magnetic HDDs and NAND flash written, which occurs at a large granularity. Unsorted are usually managed through a storage system, which is block images (UBI) is another interface for accessing organized into a few layers. The layers interact with each flash memory, which builds on top of MTD interface and other, and provide users a unified interface to access the simplifies some aspects of using raw flash memory. data being stored in the storage media. Figure 1 shows a File systems are responsible for organizing logical typical architecture for storage systems. sequences of data among the available blocks on the Zhang et al. Cybersecurity (2018) 1:1 Page 4 of 20 physical storage medium through the interface provided to the application layer), the system may simply mod- by the device driver. These include: 1) block file systems ify the metadata in the file system layer (e.g., changing built on top of block device, e.g., FAT32, EXT4, and NTFS; the block allocation table and invalidating the data being 2) flash file systems built on top of MTD device, e.g., deleted) to make the data appear to have been removed YAFFS (Robust Flash Storage: YAFFS 2002); 3) UBI file in the application layer. However, the actual content is system (Memory Technology Devices: UBIFS 2015) built still preserved in the physical storage medium layer and on top of UBI device. may be recovered by the adversary through disk foren- The highest layer is the application layer, offering an sics (Breeuwsma et al. 2007; Garfinkel and Shelat 2003). interface to the users. Therefore, secure deletion requires ensuring that the con- tent being deleted should become inaccessible at each Plausibly deniable encryption (PDE) layerofthe storagesystem. Conventional encryption is broadly used by individuals In addition, the past existence of the deleted data may and businesses to protect sensitive data. Major operat- leave artifacts in the data organization (Bajaj and Sion ing systems now increasingly support the use of full disk 2013b) or side effects on the other data (Bajaj and Sion encryption. For example, FileVault (Apple: Use filevault to 2013a). After the data have been deleted, those artifacts or encrypt the startup disk on your mac 2017)inMac OS side effects may be utilized by the adversary to learn sen- X 10.3 and later, BitLocker (Microsoft: Bitlocker 2017)in sitive information of the deleted data (Chen et al. 2016). Windows Vista and later, full-disk encryption in Android This also creates a barrier towards completely removing 4.4 and later, etc. However, traditional encryption may sensitive information. not work when data owners are captured by the adver- sary and coerced into disclosing their decryption keys. To Models and assumptions protect sensitive data against this type of coercive adver- A unified adversarial model for PDE and secure deletion saries, plausibly deniable encryption (Canetti et al. 1997; We unify the adversarial model for both PDE and secure Dürmuth and Freeman 2011;Howladerand Basu 2009; deletion by considering a snapshot adversary who can Ibrahim 2009;Klonowski et al. 2008; Meng and Wang have access to the state of a victim device. The adversary 2009; 2010; O’Neill et al. 2011) (PDE) can be utilized to is assumed to be not able to control the code of the victim hide the sensitive data by denying their very existence. Dif- system, i.e., no malicious code can be injected and hence ferent from conventional encryption, PDE encrypts origi- thecodeofthe victim system is secure.Eachaccess will nal sensitive message into ciphertexts in such a way that, allow the adversary to obtain a full snapshot of both the upon decryption, if a true key is used, the original sensi- external storage and the memory. The adversary is com- tive message can be recovered, but if a decoy key is used, a putationally bounded, and tries to illegitimately derive plausible message will be generated. When being coerced, sensitive information from the snapshots being captured. the victim can simply disclose the decoy key. Using the We consider that the adversary can have access to the decoy key, the adversary is able to decrypt the cipher- victim device once (i.e., a single-snapshot adversary)and text into the plausible message which is non-sensitive, multiple times (i.e., a multi-snapshot adversary). and is hence convinced that no sensitive information The single-snapshot adversary captures a lot of real- is stored. world scenarios. For example, an attacker steals a smart phone (Yu et al. 2014;Chang et al. 2015)oralaptop,or Secure deletion breaks into a data center obtaining a snapshot of a vic- Secure deletion is a technique designed to ensure com- tim server. The multi-snapshot adversary also captures a plete elimination of sensitive data once they become obso- lot of real-world scenarios. For example, an attacker peri- lete. It requires a guarantee that an adversary should odically breaks into a hotel room, obtaining a “memory neither recover the deleted data, nor learn anything about dump” of a victim’s smart phone; a border checker period- them. However, achieving such a guarantee in modern ically obtains snapshots from a victim’s smart device (Blass computing systems is a challenging task due to the com- et al. 2014;Petersetal. 2015). plication of the storage systems (“The architecture of a The only unique attack behavior for PDE is the adver- storage system” section). A modern storage system usually sary may coerce the data owner for the decryption keys. consists of multiple layers, and performing secure dele- This behavior is not applicable to secure deletion, because tion in one layer is usually not able to eliminate the data, no key for the deleted data will be preserved after the data since data leakage may be observed in other layers. For have been securely removed. example, given a Microsoft Word document, removing data from the document itself cannot guarantee that the The assumptions required by PDE deleted data really become inaccessible. Upon receiving a Since PDE systems usually require a few common assump- delete request issued by Microsoft Word (which belongs tions. We summarize these assumptions in the following. Zhang et al. Cybersecurity (2018) 1:1 Page 5 of 20 – The PDE software should be merged into the code layer cannot ensure that traces of the sensitive data stream of the device (e.g., part of the Android will not be observed by the snapshot adversary in the framework), such that its availability is widespread, underlying file system and physical storage medium layer, and an attacker cannot simply compromise especially when the sensitive data need to be updated deniability based on the availability of software over time and the adversary can obtain multiple snap- support. In addition, PDE requires changing a few shots (“A unified adversarial model for PDE and secure system components (e.g., booting process). The deletion” section). Therefore, when being applied to stor- adversary who can perform reverse engineering or age systems, rather than simply use encryption, two types dynamic analysis over those components will of PDE techniques, steganography and hidden volumes, unavoidably compromise deniability. The PDE are used to provide deniability. systems cannot defend against this type of adversary. The first type of PDE technique is steganography – The adversary will know the design of PDE. However, (Andersonetal. 1998). The basic idea of steganography is he/she does not have any knowledge on secret to hide sensitive data within regular file data. For exam- information (e.g., keys and passwords) being required ple, the sensitive data can be computed by performing to open/operate the PDE mode. an XOR operation over a few cover files (Anderson et al. – The adversary is rational and will stop coercing the 1998). A main concern of the steganography technique is device’s owner once he/she is convinced that the to avoid over-writing the hidden sensitive data, since they decryption keys have been revealed. are actually part of the regular data. This can be mitigated – The adversary can not capture a device working in by creating and storing (secretly) multiple copies of the the PDE mode or after a crash of the PDE mode. sensitive data, which in return will lead to inefficient use Otherwise, he/she can trivially retrieve sensitive of disk space. hidden data, compromising deniability. The other type of PDE technique is hidden volumes – The operating system, bootloader, baseband OS, (TrueCrypt: Free open source on-the-fly disk encryption firmware are all malware-free. In addition, the PDE software.version 7.1a 2012). The hidden volumes tech- mode is malware-free. nique works as follows: There are two encrypted volumes on the disk, a public volume and a hidden volume. The Protecting data confidentiality against coercive public volume is encrypted using a decoy key and the hid- adversaries via PDE den volume is encrypted using a hidden key (i.e., the true Ensuring confidentiality of the data being preserved in key). The public volume is placed on the entire disk and the hidden volume is usually placed from a secret offset personal computing devices can be achieved by encryp- towards the end of the disk (i.e., the hidden volume is part tion. However, traditional encryption cannot defend against coercive adversaries (“A unified adversarial model of the public volume, see Fig. 2). Note that initially the for PDE and secure deletion” section). Therefore, we entire disk is filled with random data and the data written mainly focus on plausibly deniable encryption (PDE), to the public volume should be placed sequentially from which can protect confidentiality of the data present in the beginning of the disk to reduce probability of over- the computing devices against both coercive and non- writing the hidden volume. When the victim is coerced coercive adversaries. into revealing the encryption key, he/she can disclose the decoy key, and the attacker will use the decoy key to PDE – from Theory to Practice decrypt the entire disk and cannot distinguish the hidden An ideal PDE would be a special encryption, which can volume from the random noise being filled initially, and is encrypt sensitive plaintexts into ciphertexts, such that the thus convinced that no sensitive data are stored. The hid- ciphertexts can be decrypted into either original sensitive den volumes technique can be viewed as a special type of plaintexts (using true key) or plausible non-sensitive steganography technique, which always hides the sensitive plaintexts (using decoy key). This is to ensure that one data in a contiguous region being placed at the end of the key can be disclosed when the data owner is coerced. disk and remains undetected by the coercive adversaries. However, such an ideal encryption is impractical for stor- Leveraging hidden volumes and steganography, vari- age systems because: First, the existing instantiation for ous PDE systems have been built to enable deniability on PDE results in a growing size of ciphertexts (Canetti et al. computing devices. All those works can be divided into 1997), which itself could be an indication of the existence two categories, PDE for desktop computers and PDE for of deniability. Second, a modern storage system is usually mobile devices. complicate, consisting of multiple layers, e.g., applica- tion layer, file system layer, physical storage medium PDE for desktop computers The existing PDE systems for desktop computers mainly layer (“The architecture of a storage system”section). rely on steganography and hidden volumes. In a few Simply encrypting the data using PDE in the application Zhang et al. Cybersecurity (2018) 1:1 Page 6 of 20 Fig. 2 The hidden volumes technique recent works, researchers also incorporate Oblivious same file system, data leakage may occur. A user of StegFS RAM (ORAM) in order to defend against multi-snapshot cannot deny the existence of hidden files, due to exis- adversaries. tence of the modified EXT2 driver and the external block table. However, the user can keep the number of security PDE using steganography. Steganographic file systems levels secret. have been initially proposed by Anderson et al. (1998), To further improve efficiency and reliability, Pang et al. with two alternative schemes. In their first scheme, the designed another steganographic file system (Pang et al. system has a number of cover files originally consisting of 2003; 2004). They use a bitmap to mark the blocks being random bits and the user’s files can be computed using used by the hidden files, and thus eliminate the need of a subset of cover files. This scheme uses a password and storing multiple copies of a file, alleviating the reliabil- the file name to determine which cover files are used. Its ity issues and I/O inefficiencies. However, the adversary drawback is requiring storing a large number of cover files. may be able to identify existence of hidden files, because Moreover, to accommodate files of arbitrary length, the the hidden blocks, being marked as used, do not pos- cover files must be relatively large. Their second scheme is sess a directory record. Three approaches are used to mitigate the aforementioned compromise. First, a few based on the computational security of block ciphers. The blocks, which do not store hidden data, are abandoned system is initialized by filling the entire hard disk with ran- dom data. A sensitive file is encrypted and stored at disk and marked as being used during initialization. Second, location being derived from the file name and the pass- when the system creates a new file, several additional word, and the encryption key is also derived in a similar blocks are allocated and filled with random noise. Third, manner. In this way, the adversary is not be able to dis- to prevent adversaries from identifying whether a block tinguish blocks containing hidden data from free blocks stores hidden data, the system maintains a few dummy being filled with random noise. However, when the disk hidden files and periodically updates them in the back- is increasingly filled by hidden files, collision of disk loca- ground. These mitigation approaches however, increase tions may increase, leading to significant over-writes. This overhead of disk space usage. is mitigated by writing each block to several disk locations. Zhou et al. further improved Pang et al.’s work by adding Inspired by the second scheme of Anderson et al. (1998), dummy transactions to obfuscate hidden files in cloud McDonald et al. proposed StegFS (McDonald and Kuhn storage (Zhou et al. 2004). Although the reliability and 2000), an EXT2-based file system which uses an exter- I/O efficiency have been improved, disk space overhead nal block allocation table to record entries for disk blocks. remains large. Troncoso et al. (2007) presented traffic StegFS supports a few security levels, each with a sep- analysis attacks on the file update algorithm proposed by arate password. To prevent overwriting the data from Zhou et al. (2004). Their attacks can detect file updates a security level which is closed, StegFS writes data in and reveal existence as well as location of files. Specif- a redundant manner. When multiple security levels are ically, they can detect files which occupy multiple data open, since hidden and regular files are present in the blocks with only two updates. Moreover, their attacks Zhang et al. Cybersecurity (2018) 1:1 Page 7 of 20 can also reveal files which occupy only one data block applications. Modern operating systems are not designed after a sufficient number of access operations. Han et al. to preserve deniability and may perform many unexpected proposed a dummy-relocatable steganographic (DRSteg) behaviors. As a result, even when the operating system file system (Han et al. 2010) to provide deniability in a runs properly, information relevant to hidden data may be multi-user environment. By sharing dummy data among leaked. For example, some operating systems (e.g., Win- multiple users in the system, DRSteg is able to increase dows) automatically create shortcuts to hidden files when the level of deniability being provided to individual users. they are used, and those shortcuts may be stored in reg- In DRSteg, in order to free dummy data without destroy- ular non-hidden directories. The shortcuts may contain ing user data, a novel dummy relocation mechanism is information about the hidden file, e.g., file name, loca- used to allow individual users to distinguish dummy data tion, length, access time, creation time and even volume from other users’ data. It can also prevent adversaries from serial number of the file system on which the hidden file distinguishing dummy and user data even after obtain- is stored. If the adversary localizes those shortcuts, it may ing multiple snapshots. There are also a few StegFS-based suspect existence of the hidden data, compromising deni- implementations including RubberhoseFS (Assange et al. ability. Another possible leakage source is the primary 2012) and Magikfs (Varun et al. 2017), which are no longer application. The primary applications are not necessarily maintained and the existing implementations may not be designed to preserve deniability, and may leak information compatible with modern Linux operating system. about the hidden data. For example, primary applica- tions may create redundant files to prevent data loss. If PDE using hidden volumes. Disk encryption tools like those files are not properly deleted, the content of the TrueCrypt (TrueCrypt: Free open source on-the-fly disk hidden data may be leaked. Finally, non-primary applica- encryption software.version 7.1a 2012)and FreeOTFE tions, such as desktop search applications, may access the (Sourceforge: FreeOTFE A free “on-the-fly” transparent files being stored in the hidden volume. Those applica- disk encryption program for PC and PDAs 2017) use hid- tions may cache snapshots of the hidden files and store den volumes to provide plausible deniability. TrueCrypt them for a later use. The adversary may also compromise supports user data encryption with several ciphers includ- deniability from those cached data. ing AES, TwoFish, Serpent, and a cascade of these ciphers in the XTS mode. The header of each TrueCrypt volume Other PDE systems for desktop computers. If the is filled with random data (e.g., salt) or encrypted with the adversary can capture multiple snapshots of hidden vol- header key which is derived from the corresponding pass- ume at different points of time, they can detect exis- tence of the hidden volume, by simply comparing dif- word using PBKDF2. Therefore, the entire volume appears ferent snapshots and identifying whether “free” blocks as being filled with randomness. When TrueCrypt loads have been changed. Therefore, TrueCrypt cannot pro- a volume, since it does not store the cipher specifica- vide deniability when facing a multi-snapshot adversary. tion, all supported ciphers will be tried with a header key Blass et al. proposed HIVE (Blass et al. 2014) to allow the (being derived from the user’s password using PBKDF2) user to deny existence of the hidden volume when fac- until it decrypts the volume and obtains the ASCII string ing a multi-snapshot attack. This is achieved by hiding “TRUE” from a certain block in the header. Then True- Crypt decrypts the encrypted master volume key using every access of the disk using Oblivious RAM (ORAM) the header key. Note that the master volume key is gener- (Goldreich and Ostrovsky 1996;Stefanovetal. 2013), ated randomly upon creating the volume. If a TrueCrypt which was originally designed to hide patterns of access to hidden volume is created, there will be also a hidden remote storage. However, ORAM is known as expensive header, which contains offset of the hidden volume. The in terms of both computation and I/O. Although HIVE hidden header is tested before the public header when uses a less expensive write-only ORAM that only supports mounting a volume. TrueCrypt can also create a hidden write operations, its performance is still far from being OS in a hidden volume by creating a new partition and practical. copying the current OS to the hidden volume. When the To improve performance of HIVE, Chakraborti et al. system is booted into the hidden OS, all unencrypted vol- proposed DataLair (Chakraborti et al. 2016; 2017). Hav- umes and non-hidden encrypted volumes are mounted ing observed that revealing access patterns to the public in a read-only manner, ensuring that any OS/application- data is unnecessary, DataLair only protects operations specific leakage stays within the hidden volume. on the hidden data and ensures that they are indistin- Czeskis et al. (2008) analyzed TrueCrypt and proposed guishable from operations on the public data. In addition, three types of attacks against it. They consider three DataLair optimizes the oblivious access mechanism being types of leakage sources: 1) the operating system; and deployed for hidden data. Compared to HIVE, DataLair is twoordersofmagnitude faster in termsofpublicdata 2) the primary applications (i.e., an application that is access, and five times faster in terms of hidden data access. used to manage hidden data); and 3) the non-primary Zhang et al. Cybersecurity (2018) 1:1 Page 8 of 20 Roche et al. designed DetWoORAM (Roche et al. 2017), derives the hidden volume key; and 2) decoy password an entirely new technique for write-only ORAM, which that derives only the decoy volume key; and 3) deletion uses a deterministic and sequential writing pattern, elim- password that derives the decoy volume key and over- inating the need of any “stashing” of blocks in local state. writes the hidden volume key. When coerced, the victim In DetWoORAM, since the write will always succeed and can fake compliance, and enter the deletion password, and occur in a free block, the notion of stash can be com- then can prove to the attacker that Gracewipe has been pletely removed. They also pointed out that the construc- executed and the real key is no longer available. tion of DataLair does not satisfy write-only obliviousness, Table 1 summarizes the existing PDE systems for desk- since the process of finding free blocks leaks information top computers, being incorporated into different lay- about which blocks are free, and the adversary can tell ers of the storage systems. To defend against snapshot whether recent writes have been performed on the same adversaries, they may rely on hidden volumes, steganog- address or not. raphy, or ORAM, and provide one or multiple denia- Zuck et al. presented Ever-Changing Disk (ECD) (Zuck bility levels with different overheads. Schemes based on et al. 2017) to achieve deniable storage system. Their steganography (e.g., McDonald and Kuhn 2000;Pang design follows three requirements: 1) resistance to multi- et al. 2003) suffer from high performance overhead since snapshot adversaries; and 2) ensuring that hidden data will they write multiple copies of files or dummy writes not be destroyed when a user is writing the public volume; for steganography. Hidden volume mechanism (True- and 3) using normal system operations on public data to Crypt: Free open source on-the-fly disk encryption soft- disguise writes to hidden data. In ECD, the storage space is ware.version 7.1a 2012) introduces low overhead since the separated into two parts: a part containing the public data hidden files are stored in the free space and the over- volume and the other containing the hidden data volume. write problem is mitigated by linear space allocation. The hidden volume is visible to the system when the user Since ORAM is known to be expensive, schemes based enters the secret key. ECD uses a large volume of pseu- on ORAM (e.g., Blass et al. 2014; Chakraborti et al. 2017; dorandom data to hide the sensitive data. A portion of Roche et al. 2017) unavoidably have high overheads. data from the volume are periodically migrated using nor- mal firmware operations to obfuscate writes to the hidden PDE for mobile devices data. Since hidden and pseudorandom data blocks are Compared to desktop computers, mobile devices are usu- constantly relocated and modified, the hidden data may ally different in two aspects: First, they are equipped with eventually be overwritten without knowing the secret key. less computational power. Second, they usually use flash To mitigate the overwrite issue, the rate of internal data memory (“Flash memory”section)asstorage.The exist- migration is controlled by the user and the user should ing PDE systems for mobile devices can be divided into enter the secret key periodically. two categories: 1) The PDE systems built on top of block Zhao et al. proposed Gracewipe (Zhao and Mannan devices. This type of PDE systems views flash memory as a 2015), by which the victim can provably destroy/erase data black box, which exposes a block-access interface through when being coerced, hoping that a reasonable adversary FTL (“The architecture of a storage system”section). will find no reason to keep holding him/her. Gracewipe 2) The PDE systems built on top of flash memory. This works as follows: During setup, the user selects three pass- type of PDE systems directly work on top of flash memory words, which can be used to derive the key for encrypting to provide deniability while handling the special nature of the corresponding volume: 1) hidden password that only flash. Table 1 Summary of PDE systems for desktop computers Scheme Method Adversary Deniable level Layer Scheme 1 (Anderson et al. 1998) Steganography Multi-snapshot Arbitrary File system Scheme 2 (Anderson et al. 1998) Steganography Single-snapshot Arbitrary File system StegFS (McDonald and Kuhn 2000) Steganography Single-snapshot Arbitrary File system StegFS (Pang et al. 2003; 2004) Steganography Multi-snapshot Arbitrary File system StegHide (Zhou et al. 2004) Steganography Multi-snapshot Arbitrary File system TrueCrypt (2012) Hidden volume Single-snapshot One Block device HIVE (Blass et al. 2014) ORAM Multi-snapshot Multiple Block device DataLair (Chakraborti et al. 2017) ORAM Multi-snapshot Multiple Block device DetWoORAM (Roche et al. 2017) ORAM Multi-snapshot Multiple Block device ECD (Zuck et al. 2017) Hidden volume Multi-snapshot One Block device Zhang et al. Cybersecurity (2018) 1:1 Page 9 of 20 PDE systems on top of block devices. Skillen et al. fast mode switching, MobiHydra introduces a special par- (Skillen and Mannan 2013, 2014; Skillen 2013)designed tition called shelter volume on the external storage, which Mobiflage, the first PDE scheme for mobile devices. They is used as a temporary storage partition which can tem- provided two versions of implementations: one in external porarily store the sensitive data being created in the pub- storage for FAT32 file system (Skillen and Mannan 2013), lic mode, without the need of entering the PDE mode and the other in internal storage for modified EXT4 file for storing hidden sensitive data. The data stored in the system (Skillen and Mannan 2014). The main contribu- shelter volume will be immediately synchronized to the tion of Mobiflage lies in its first incorporation of hidden hidden volume when the hidden mode is entered, and volumes technique to the Android devices. It works as are then eliminated from the shelter volume. To avoid follows: First, it fills the external storage with random deniability compromise, the sensitive data stored in the data. Second, it creates two volumes, a public volume for shelter volume will be encrypted by a random key which storing public non-sensitive data, applications and set- is encrypted by the public key and stored in the shelter tings, and a hidden volume for storing sensitive data. volume. In addition, some dummy files are maintained in Correspondingly, there are two operation modes, a pub- the shelter volume and updated periodically to obfuscate lic mode and a PDE mode. The public mode is used to the writes of hidden sensitive data to the shelter volume. manage the public volume for daily use, providing stor- MobiHydra, however, cannot eliminate the assumption of age encryption without deniability. In this mode, the data requiring a physical or an emulated FAT32 SD card. are encrypted with a decoy key, which is derived from a To eliminate the limitations of Mobiflage and MobiHy- decoy password. The user will be asked to provide the dra, Chang et al. designed MobiPluto (Chang et al. 2015), decoy password during the booting in order to enter the a file system friendly PDE design. The basic idea of Mobi- public mode. The PDE mode is used to manage the hid- Pluto is introducing an additional software layer between den volume which stores sensitive data whose existence the PDE and the file system. This software layer should needs to be denied when being coerced. In this mode, the satisfy three requirements: 1) Its existence should not be data are encrypted with a true key, which is derived from an indication of PDE; 2) It should provide virtual vol- atruepassword. Theusershould provide thetruepass- umes to file systems, and any block-based file systems word during system boot to activate the PDE mode. The can be deployed on a virtual volume; 3) It should convert exact location of the hidden volume is derived from the non-sequential allocation from a file system to sequential true password. allocation in the underlying PDE. To build such a layer, The FAT32 version of Mobiflage (Skillen and Mannan MobiPluto uses thin provisioning (Thornber), because: 2013) is specially designed for external storage formatted First, thin provisioning has been implemented by dm- using FAT32 file system, which requires the support of a thin-pool module, which has been a well-established tool physical or an emulated FAT32 SD card. This is because, in Linux kernel; Second, thin provisioning can allow to the hidden volume is part of the public volume, and is create thin volumes, each of which can be used to deploy placed at the end of the disk. In other words, if an EXT- any block-based file systems. Third, thin provisioning can like file system is deployed for the public volume, the data transform non-sequential allocation on the thin volume to stored in the hidden volume may be easily overwritten sequential allocation on the underlying storage. By com- by the public data due to the nature of EXT file sys- bining thin provisioning and hidden volumes, MobiPluto tem’s random allocation. To eliminate the aforementioned is able to achieve a “file system friendly” PDE design. assumption, the EXT4 version of Mobiflage (Skillen and Chang et al. further improved the usability of MobiPluto Mannan 2014) modifies the EXT4 driver such that it can in their extended work (Chang et al. 2018)byintroduc- support a sequential inode allocator and can be deployed ing a fast switching mechanism and using NFC cards to for the public volume to avoid overwriting the hidden vol- store strong PDE passwords. For PDE systems on mobile ume. However, the modification of the EXT4 driver itself devices, fast switching is a desired feature. When a device may be an indication of the existence of PDE. owner faces an emergency and wants to collect sensi- Yu et al. proposed MobiHydra (Yu et al. 2014)to tive information, he/she needs to instantly switch the improved Mobiflage in three aspects: 1) It can mitigate device to the hidden mode. However, it needs more than a novel booting-time attack being faced by Mobiflage; 1 min for the prior mobile PDE systems to switch modes, 2) It can support multiple levels of deniability. 3) It sup- because a full device rebooting is usually required. Fast ports mode switching without rebooting. To mitigate switching mechanism eliminates the need for rebooting the booting-time attack, MobiHydra obfuscates the time the device and the switching time is reduced to less than required for a wrong password (i.e., an arbitrary password 10 s (Chang et al. 2018). Their idea is to restart only except the true and the decoy password) during booting the Android framework rather than the entire device, such that the adversary is not able to identify the existence significantly reducing the switching time. For the hidden of PDE by simply entering a wrong password. To support volume, a strong password is required to protect sensitive Zhang et al. Cybersecurity (2018) 1:1 Page 10 of 20 data. However, users tend to choose weak passwords, thus properties in YAFFS2 to provide deniability. Therefore, it security may not be ensured. To address this issue, they is incompatible with the flash-based block devices using proposed to use NFC cards to store strong passwords for FTL, the most popular form of flash storage being used the user. Their observation is that most of the modern in mobile devices nowadays. Second, it suffers from deni- mobile devices are equipped with NFC features. ability compromises (Jia et al. 2017). This is because, to MobiCeal (Chang et al. 2018) designed the first block- prevent data at lower security levels from overwriting the based PDE system for mobile devices which can defend data at higher security levels, it disables garbage collection against a multi-snapshot adversary. The fundamental idea at the lower security levels. The adversary can easily iden- of MobiCeal is to use dummy writes to obfuscate writes tify this abnormal behavior and suspect existence of PDE performed in hidden mode. In this way, even though the (Jia et al. 2017). adversary can have access to multiple snapshots of the DEFTL (Jia et al. 2017) incorporated PDE into FTL. block image, it cannot tell whether a write is issued by DEFTL also has two modes, a public and a hidden mode. the hidden mode or not. In addition, MobiCeal can sup- The deniability of DEFTL is achieved by using the data port multiple deniability levels. Finally, it identifies various (and theirbehavior) in thepublicmodetodenythe side channel leakage present in prior PDE schemes for data (and their behavior) in the hidden mode. Most mobile device and eliminates them. importantly, to prevent the data written in the public mode from over-writing the data written in the hidden PDE systems on top of flash memory. Flash memory mode, DEFTL carefully modifies the block allocation and has significantly different nature compared to mechanical garbage collection strategies in the FTL such that the two disks, e.g., flash memory is update unfriendly and vulner- modes can be “stealthily” isolated without being known able to wear (“Flash memory” section). All the aforemen- by the adversary. Specifically, the public volume will allo- tioned block-based PDEs unfortunately may suffer from cate flash blocks from the head of the block pool and the deniability compromises in the underlying flash storage hidden volume will allocate flash blocks from the tail of due to the handling of the unique nature of flash memory. the pool. In addition, garbage collection in the two modes This is because: The unique characteristics of flash mem- will be modified as: In the public mode, garbage collec- ory require a special internal management, which creates tion will be performed actively to fill the head of the pool; a different view of data in flash memory, independent of in the hidden mode, garbage collection will be performed the view on the block layer. By having access to the raw actively to fill the tail of the pool. This can avoid that flash, the adversary can obtain this different view, which the public mode has used all the blocks in the head and starts to use the blocks in the tail, over-writing the hidden may allow it to observe those unexpected “traces” of sensi- sensitive data. DEFTL also provides a few attacks on the tive data, whose existence needs to be denied. To eliminate existing PDE systems for mobile devices. the aforementioned deniability compromise, a few PDE Table 2 summarizes the existing PDE systems for mobile systems directly incorporate PDE into flash memory. devices. They either use hidden volumes or steganogra- Peters et al. (2015)introducedDEFY, adeniable phy, and provide one or multiple deniability levels when encrypted file system based on flash file system YAFFS2 facing single-snapshot or multi-snapshot adversaries. (Robust Flash Storage: YAFFS 2002). In DEFY, operations Schemes based on hidden volumes (Skillen and Mannan on a higher security level are indistinguishable from the 2013, 2014;Yuetal. 2014;Chang et al. 2015, 2018) operations on a lower security level. In addition, DEFY usually have low performance overhead, but they can can mitigate over-writes of hidden data in the higher secu- only defend against single-snapshot adversaries. DEFY rity level by taking advantage of special properties offered by a log-structured file system. DEFY however, suffers (Peters et al. 2015) relies on steganography to defend from a few limitations. First, it strongly relies on system against multi-snapshot adversaries, but the performance Table 2 Comparison of PDE systems for mobile devices Scheme Method Adversary Deniability level Layer Mobiflage-FAT32 (Skillen and Mannan 2013) Hidden volume Single-snapshot One Block device Mobiflage-EXT4 (Skillen and Mannan 2014) Hidden volume Single-snapshot One Block device MobiHydra (Yu et al. 2014) Hidden volume Single-snapshot Multiple Block device MobiPluto Chang et al. (2015, 2018) Hidden volume Single-snapshot Multiple Block device MobiCeal (Chang et al. 2018) Steganography Multi-snapshot Multiple Block device DEFY (Peters et al. 2015) Steganography Multi-snapshot Arbitrary File system DEFTL (Jia et al. 2017) Hidden volume Single-snapshot One Controller (FTL) Zhang et al. Cybersecurity (2018) 1:1 Page 11 of 20 overhead is high. MobiCeal (Chang et al. 2018)usesa However, once sensitive data are discarded, the data lightweight “dummy write” mechanism to defend against owner may want to permanently remove them. Protec- multi-snapshot with acceptable performance overhead. tion of data confidentiality requires securely disposing those data to prevent the adversary from recovering all or Discussions portion of them. This is achieved by using secure deletion. About randomness being used in hidden volumes Hard disk drives (HDDs) and NAND flash mem- ory dominate the storage media of personal computing technique. The idea of hidden volumes technique is to hide the encrypted hidden volume among the random- devices. However, they are completely different in nature, ness being filled initially. To make sure that the ciphertexts and hence different secure deletion approaches should be being generated are indistinguishable from random bits, used to eliminate data from them. In the following, we the random bits can be drawn from the same distribution summarize secure deletion approaches for HDDs-based as the ciphertext space by using the encryption function and flash memory-based storage systems, respectively. itself as the PRNG (Skillen and Mannan 2014). This can We also summarize the recent works which complement ensure that the ciphertexts and the randomness are from conventional secure deletion approaches by taking care of the same source. In addition, the encryption being used the impacts created by past existence of the deleted data. in the hidden volumes technique is full disk encryption (FDE), which treats each disk sector as an autonomous Secure deletion for HDDs-based storage systems unit and assigns sector-specific IVs for chaining modes An HDD is a magnetic medium which supports in- such as CBC and XTS (Skillen and Mannan 2014). This place updates. Therefore, in HDDs-based storage sys- can help eliminate the correlations among ciphertexts due tems, when a file block is updated (or deleted), its old to similarities in files (e.g., file heads). version can be simply replaced by its new version (or random noise) in the storage medium to achieve secure Transferring data between desktop computers and deletion. mobile devices. SincePDE canbedeployedineither Physical medium layer. Due to lack of semantics of file desktop computers or mobile devices, there is a possibility system, single-file sanitization is not feasible at the phys- that the data owner would like to exchange data (e.g., files) ical medium layer. To delete the data, a naivest way is to between these two different platforms. The data exchange overwrite/destroy all the data on the physical medium. should be conducted in such a way that deniability should Tools such as degaussers can be used to sanitize data on not be compromised. When the file is non-sensitive, it HDDs (Kisseletal. 2006). can be simply transferred from a desktop computer to a Controller layer. At the controller layer, there are sev- eral standardized interfaces that permit reading/writing mobile device (or vice versa) without any deniability con- of fixed-sized blocks. Similarly, there are no semantics of cern. This can be done by simply using Internet or direct file system in this layer, thus the controller must sani- copying via cable, when both the source device and the tize every block to achieve secure deletion. To delete the destination device are working in the public mode. When the file is sensitive whose existence needs to be denied, to data, sanitize commands and overwrite techniques are transfer it from a desktop computer to a mobile device (or widely used in HDDs, e.g., secure erase commands offered vice versa), there are two known options: 1) If the file is by both SCSI (Incits: Scsi storage interfaces 2016)and transferred via Internet, to allow the data owner to deny ATA (Team work systems: Advanced technology attach- the transferring of this file (the attacker can collude with ment 2017). These sanitization commands work like a Internet ISPs and identify this event), covert communica- button that erases all data on the device by exhaustively tions approaches (Frèche et al. 2017;Huetal. 2017)can be overwriting every block with zeros or ones. used. 2) If the file is transferred via direct copying, the file Block device layer. Reardon et al. (2013)proposeda canbereadfromthe source device andwrittento thedes- secure deletion approach targeting persistent storage. tination device, when both devices are working the hidden Their approach relies on encryption and key wrapping. mode. Since the attacker cannot capture the device work- They useakey disclosure graphtomodel theadversarial ing in the hidden mode (“The assumptions required by knowledge about key generation and wrapping history. In PDE” section), the deniability will not be compromised. addition, a small securely-deleting key-value map is used Once the file is successfully transferred, it will be pro- to discard encryption key of the data, achieving secure tected by the PDE system in the destination device. deletion. File system layer. When deleting files in the EXT2, there Ensuring confidentiality of the deleted data via is a sensitive attribute for files and directories to indicate secure deletion that secure deletion should be used. (Bauer and Priyantha 2001) provided a patch that implements this attribute. By PDE is used to ensure confidentiality of sensitive data marking a block as free, the patch passes the free block which are preserved in the personal computing devices. Zhang et al. Cybersecurity (2018) 1:1 Page 12 of 20 to a kernel daemon, which maintains a list of blocks that deletion interface, which can be used to overwrite data must be sanitized. If the free block stores data from a with zeros in the underlying file system. For MySQL, sensitive file, instead of returning it to the file system (Stahlberg et al. 2007)proposedanapproach to delete as an empty block, this free block will be added to the entries by overwriting them with zeros, and the trans- work queue. When the system is idle, the sanitization dae- action log is encrypted and can be securely disposed by mon runs asynchronously to perform sanitization over the deleting the encryption key. For SQLite (SQLite: Pragma work queue, allowing the user to achieve immediate file statements 2017), there is a compile-time option to enable deletion. a secure deletion feature that overwrites deleted records (Joukov and Zadok 2005)proposedafile system with zeros. extension, purgefs, which uses block-based overwrit- Table 3 summarizes the existing secure deletion ing when blocks are returned to the file system’s free approaches for HDDs-based storage systems. Those blocks list. It supports overwriting file data and metadata approaches are incorporated into different layers of a stor- for all files or just files marked as sensitive. (Joukov age system and rely on either overwriting or encryption and Zadok 2006) also proposed three secure deletion for secure deletion. They may also have different dele- approaches for EXT3. The first approach is called EXT3 tion granularity. Generally, the encryption-based secure basic, which securely deletes data (but not metadata) deletion approaches for HDDs-based storage systems once by overwriting it. The second approach is called (Reardon et al. 2013;Petersonetal. 2005;Zhaoand EXT3 comprehensive, which overwrites file data and Mannan 2015) have relatively low performance overheads metadata by a configurable overwriting scheme. Both since only keys need to be deleted to achieve secure the aforementioned two overwriting-based approaches deletion. can securely delete all the data or just files whose extended attributes include a sensitive flag. The third Secure deletion for flash memory-based storage systems approach is based on intercepting files deletion events, The aforementioned secure deletion approaches for i.e., unlinking and truncating a file. The file to be deleted HDDs-based storage systems rely on properties of hard is moved into a special secure deletion directory, and disks: magnetism-based and supporting in-place updates. a background user-level tool shred (Plumb 2010)will However, flash memory does not possess these proper- delete the files in the secure deletion directory at regular ties. Wei et al. (2011) performed a series of experiments intervals. to show that the deletion techniques that work well for Peterson et al. (2005) optimized secure deletion for HDDs may not work properly for NAND flash. As NAND versioning file systems using an all-or-nothing transform flash is not magnetism-based, the degaussing method may (AONT). Using AONT, each data block is extended into damage NAND flash chips and render data unreadable, an encrypted data block along with a small stub. If any part but all the data may still remain intact (Wei et al. 2011). of the ciphertext is deleted, the entire message can not be Wei et al. (2011) created 1000 small files on an SSD, decrypted any more. In this way, a specific version of a file then dismantled the drive, and searched for the content can be quickly deleted by simply overwriting all the stubs. of those files. They found that the SSD contained up to In addition, to delete a large log file to which data have 16 stale copies of the tested files. This is because the been appended only, securely deleting all the blocks in its FTL creates redundant file copies during garbage collec- most recent version will achieve secure deletion on all its tion and out-of-place updates, which unfortunately will past versions. complicate the secure deletion design for flash memory. Application layer. The application layer can only inter- They also tested 13 single-file overwriting-based saniti- act with file system through a POSIX-compliant interface. zation tools (LSoft Technologies Inc: Active@ KillDisk A user-level application can securely erase all the data 2017; GEEP EDS LLC: Darik’s Boot and Nuke 2017), to on the storage medium by invoking a secure erase com- find out whether they work correctly for flash memory. mand (Incits: Scsi storage interfaces 2016;Teamwork Unfortunately, all those tools were not able to sanitize data systems: Advanced technology attachment 2017)inthe from flash memory: between 4 and 75% of the files’ con- hardware controller’ interface. A few files overwriting tent remained in the SATA SSDs, and between 0.57 and tools, e.g, srm (Jagdmann 2015)and wipe (Durak 2006), 84.9% of the data remained in USB drives. All the afore- can be used to securely remove files. Gracewipe (Zhao mentioned results indicate that securely deleting data and Mannan 2015), as has been discussed in “PDE for from NAND flash is much more challenging compared desktop computers” section, can achieve secure and veri- to HDDs, due to the special nature of NAND flash. In fiable deletion of encryption keys through a special dele- the following, we summarize the existing works aiming to tion password by taking advantage of TPM and Intel securely remove data from flash-based storage systems, TXT, thus making the encrypted data permanently inac- which are divided into two categories: overwriting-based cessible. In the case of database, there is also a secure and encryption-based. Zhang et al. Cybersecurity (2018) 1:1 Page 13 of 20 Table 3 Comparison of features offered by different secure deletion approaches for HDDs-based storage systems Scheme Method Layer Deletion granularity Degaussers (Kissel et al. 2006) Degaussing Physical medium The entire storage medium Secure erase commands (2016, 2017) Overwriting-based Controller The entire storage medium Reardon et al. (Reardon et al. 2013) Encryption-based Block device Single file Bauer et al. (Bauer and Priyantha 2001) Overwriting-based File system (EXT2) Single file Purgefs (Joukov and Zadok 2005) Overwriting-based File system (EXT2) Single file Joukov et al. (Joukov et al. 2006) Overwriting-based File system (EXT3) Single file Peterson et al. (Peterson et al. 2005) Encryption-based File system Single file File overwriting tools (Jagdmann 2015;Durak 2006) Overwriting-based Application Single file Gracewipe (Zhao and Mannan 2015) Encryption-based Application The entire storage medium Stahlberg et al. (Stahlberg et al. 2007) Based on overwriting Application (for database) Single entry/record and SQLite (SQLite: Pragma statements 2017) and encryption Overwriting-based secure deletion for NAND flash. some others, it does not cause any errors. They intro- A common idea for the overwriting-based secure deletion duced scrub budget, which refers to the number of times approaches is to replace the deleted data with meaningless that the NAND flash can allow to be scrubbed with- information, e.g., noisy random data. out exhibiting a significant risk of data errors. When the Physical medium layer. Similar to the physical medium scrub budget for a block is exceeded, secure deletion will layer in the HDDs-based storage systems (“Secure dele- be instead performed by other approaches (e.g., invok- tion for HDDs-based storage systems”section), data can ing garbage collection). More recently, Qin et al. (2013) notbedeleted in this layerdue to lack of semanticsof incorporated RAID-5 architecture to enhance the reliabil- the file system. To realize secure deletion in this layer, the ity and eliminate the negative effect of reprogram on flash entire flash chip should be destroyed. memory. Controller layer. A main type of flash controller is using File system layer. Sun et al. (2008) proposed a secure FTL to handle the special nature of NAND flash and to deletion method in YAFFS by investigating characteris- provide a block access interface to upper layers (“The tics of NAND flash memory. They proposed two secure architecture of a storage system” section). To securely deletion approaches, zero overwrite (similar to the scrub- delete data, the simplest way is to erase the correspond- bing (Wei et al. 2011)) and block erase. Especially, they ing flash blocks in the FTL (i.e., block erasure). However, define a costs-benefits model by comparing the overwrite erasures can only be performed in terms of flash blocks cost on the deleted pages and the erase cost on the block (“Flash memory” section). This will be overkill if only a that contains the deleted pages. Additionally, a new adap- portion of data being stored in a flash block needs to be tive hybrid scheme is applied to select the cheaper one deleted. Considering content of a file may be distributed between the two secure deletion approaches. Another in different pages of different flash blocks, sanitizing a file kernel-level zero-overwriting secure deletion approach using block erasure will be unavoidably expensive. was also proposed by Reardon et al (2012). Wei et al. proposed scrubbing (Wei et al. 2011)to By adding a new communication channel between the address the aforementioned issue. As programming indi- file system and the device driver, the file system can vidual pages is possible, the idea of scrubbing is to re- inform the device that particular blocks are no longer program the page, where the data should be securely valid, e.g., Trim (Intel Corporation: Intel Solid-State Drive deleted, to turn all its remaining ‘1’ bits into ‘0’. Note that Optimizer 2009) command and TrueErase (Diesburg et al. flash memory allows to individually program bit ‘1’ to ‘0’, 2012). With the information of invalid blocks, the device but the reverse operation is not feasible except perform- driver can implement its own efficient secure deletion ing a block erasure. A major concern of scrubbing is that without requiring data blocks to be explicitly overwrit- it may result in undefined behaviors due to possibility of ten by the file system. Especially, TrueErase is designed introducing read errors. To handle this concern, Wei et al. for the blocks belonging to files specifically marked as examined error rates for different types of flash memory sensitive. and showed that the error rates vary widely. For some Application layer. Since the application layer cannot flash devices, scrubbing causes frequent errors, while for directly touch the lower layers, secure deletion can only Zhang et al. Cybersecurity (2018) 1:1 Page 14 of 20 be achieved by filling the entire remaining free space. approach to perform standard data sanitization which can Reardon et al. (2012) proposed two user-level filling-based satisfy government agencies’ requirements (NSA/CSS and secure deletion approaches for YAFFS: purging and bal- DoD 5220.22-M) for the secure deletion. looning. Their basic idea is to fill the entire free space of Reardon et al. (2012)proposeddatanodeencrypted file the file system, such that all unused blocks of the physi- system (DNEFS), which enables secure data deletion for cal medium will no longer contain sensitive information. flash memory. They also incorporate DNEFS into flash By completely filling the file system’s empty space with file system UBIFS (Memory Technology Devices: UBIFS noise, all previously data deleted by the user are guaran- 2015). In DNEFS, they divide the entire flash memory into teed to have been erased. Compared to purging which two areas: a small key storage area and a large main data ensures rapid secure deletion of data from user-space, bal- storage area. They encrypt each data node (i.e., the unit of looning achieves a probabilistic continuous secure dele- I/O) with a unique key, and collocate the keys in the key tion guarantee by reducing the block reallocation period. area. Secure deletion is achieved by removing keys, which Braga et al. (2014) proposed two user-level approaches to can be performed efficiently, as keys are condensed in a securely delete files on Android smart phones. The one small area. that is designed to delete unencrypted files is also based DEFY (Peters et al. 2015) also provides secure dele- on filling. The cost of the filling-based secure deletion tion, complementary to its deniability. It leverages all- approaches is proportional to the size of free space avail- or-nothing transform (AONT), a cryptographic function able on the physical medium. A larger size of free space which can ensure that a missing portion of a message will will lead to a higher overhead in order to fill it (Reardon render the entire message irrecoverable. In this way, DEFY et al. 2013). The efficiency can be improved by perpetually can efficiently achieve secure deletion by only removing maintaining the free space of the physical medium within a small portion of the data being deleted. Braga et al. a limited range (Reardon et al. 2012). (Braga and Colito 2014) proposed two secure deletion approaches for Android phones, one of which is based on Encryption-based secure deletion for NAND flash. encryption. They modified the key management of EncFS Data can be rendered inaccessible by encrypting them (Wang et al. 2012), an encrypted file system, to ensure that and deleting the corresponding key. Boneh et al. pro- every file is encrypted with a unique key and a random posed the first encryption-based solution that securely IV. The removal of the unique key and IV makes the file deletes encrypted data stored on the tape by deleting the irrecoverable. cryptographic keys (Boneh and Lipton 1996). Table 4 summarizes the features offered by different secure deletion approaches for flash-based storage sys- Controller layer. Reliably destroying keys is challenging, tems, including overwriting-based and encryption-based as side-channel attacks based on semiconductor memory approaches. These approaches may also have differ- data remnants (Halderman et al. 2009) may allow an ent deletion granularity. Generally, the scrubbing-based attacker to recover the key or key-related information. secure deletion approaches (Wei et al. 2011;Qinet al. Swanson et al. (2010)proposedscrambleand finally erase (SAFE), which combines encryption and erasure tech- 2013;Sun et al. 2008; Reardon et al. 2012), which only niques to provide almost instant secure deletion with ver- targets at the invalid data, and the encryption-based ifiability. SAFE relies on the assumptions that data in the secure deletion approaches (Lee et al. 2008, 2010, 2011; SSDs are stored encrypted and the SSDs implement best Reardon et al. 2012), which only need to delete keys, practices of key management (e.g., the keys should never have relatively low performance overheads compared to leave the controller). It works as follows: Upon receiving the overwriting-based approaches. A major advantage of a sanitize command, it erases the controller’s key mem- encryption-based approaches is that, deleting a small key ory, such that the driver is not able to encrypt/decrypt the usually can be much more efficiently achieved compared data. It then erases every block on the device, and writes to deleting data which are large in size. all the pages with a known pattern, and erases them again. Finally, it reinitializes the device and performs a low-level A new direction for secure deletion format operation on the drive, and provides a new key to Secure deletion is used to securely dispose data once they the controller. become obsolete. This requires a security deletion guaran- File system layer. Lee et al. (2008, 2010)proposeda tee that the adversary should neither recover the deleted secure deletion approach for YAFFS, a log structured data, nor learn anything about them (“Secure deletion” NAND flash file system. By modifying YAFFS, they section). The question is, can we achieve the secure dele- encrypt files and force all keys of a specific file to be tion guarantee by simply deleting the data themselves? stored in the same block. Therefore, only one erase oper- The answer is unfortunately no. Conventional secure dele- tion approaches rely on either overwriting or encryption ation needs to be performed in order to securely delete a to make the deleted data inaccessible. However, the past file. Lee et al. (2011) then extended the aforementioned Zhang et al. Cybersecurity (2018) 1:1 Page 15 of 20 Table 4 Comparison of features offered by secure deletion approaches for flash-based storage systems Scheme Method Layer Deletion granularity Scrubbing (Wei et al. 2011) and SmSD (Qin et al. 2013) Overwriting-based Controller (FTL) A physical page Sun et al. (Sun et al. 2008) Overwriting-based File system (YAFFS) Single file Zero overwriting (Reardon et al. 2012) Overwriting-based File system (YAFFS) Single file Trim (Intel Corporation: Intel Solid-State Drive Optimizer 2009)and TrueErase (Diesburg et al. 2012) Overwriting-based File system Single file to device driver Purging and ballooning (Reardon et al. 2012) Overwriting-based Application All the invalid data Braga et al. (Braga and Colito 2014) for Overwriting-based Application All the invalid data unencrypted files SAFE (Swanson and Wei 2010) Based on overwriting Controller The entire and encryption storage medium Lee et al. (2008, 2010) Encryption-based File system (YAFFS) Single file Lee et al. (2011) Encryption-based File system (YAFFS) Single file DNEFS (Reardon et al. 2012) Encryption-based File system (UBIFS) Single file DEFY (Peters et al. 2015) Encryption-based File system (YAFFS) Single file Braga and Colito (2014) for Encryption-based File system (EncFS) Single file encrypted files existence of the deleted data may leave artifacts in the lay- structures. This also indicates that, although node 2 have out at all layers of a computing system (Bajaj and Sion been deleted, its structural artifacts remain in the data 2013b;Chenand Sion 2015) or create side effects on organization. Therefore by having access to tree T ,the the other data which have not been deleted (Bajaj and adversary may suspect the past existence of the sensitive Sion 2013a), and the adversary can potentially take advan- data which have been deleted, and tries to partially or fully tage of those structural artifacts or side effects to learn recover them. sensitive information about the deleted data (Bajaj and We also describe some concrete attack scenarios which Sion 2013a, b;Chenand Sion 2015). can take advantage of structural artifacts (Jia et al. 2016; To justify the impact of the past existence of the deleted Chen et al. 2016). Commodity NAND flash-based block data, we use a balanced binary search tree (BST) (Chen devices usually adopt a log-structured writing technique, et al. 2016). We first create a balanced BST by inserting in which flash blocks as well as pages within a block five nodes in the order of 2, 11, 13, 14, 1, and obtain tree are allocated sequentially (Min et al. 2012). As shown in T (Fig. 3a). We then delete node 2, obtaining tree T Fig. 4,A,B,C,and Dare thedatabeing writtentoNAND 1 2 (Fig. 3b). However, if we directly create the balanced BST flash, and each occupies a flash page. To securely delete by inserting nodes in the order of 11, 13, 14, 1, we will C, the user has two options: 1) We perform a scrubbing obtain tree T (Fig. 3c). This example indicates that, due (Wei et al. 2011) over the corresponding flash page. The to the past existence of node 2, T and T have different scrubbing technique however, will convert this page to a 2 3 Fig. 3 An example from (Chen et al. 2016) showing why structural artifacts matter. T , T ,and T are balanced BSTs. a Tree T , b Tree T ,and c Tree T 1 2 3 1 2 3 Zhang et al. Cybersecurity (2018) 1:1 Page 16 of 20 Fig. 4 Scrubbing-based secure deletion page with all “0” bits (i.e., a zero page). By having access flash with a full scrubbing (Wei et al. 2011)orablockera- to the storage state after deletion of C, the adversary will sure may provide the adversary a clue that there was a notice the zero page and suspect that a past deletion has deletion in the past, Jia et al. proposed to perform a par- been performed on it. 2) We encrypt A, B, C, and D tial scrubbing, i.e., partial page reprogramming and partial with different keys, and delete key for C (Reardon et al. block erasure, to only modify a portion of the bits in the 2012). However, the adversary can also find out that C has page/block storing the deleted data, avoiding producing a been deleted since it cannot be successfully decrypted to zero page or an all-“1” block. plaintexts which are semantically meaningful. A further Truly secure deletion. NFPS (Jia et al. 2016)aimed to attack will be performed to fully or partially recover C conceal the past existence of the deleted data in the by taking advantage of the correlation between C and its NAND flash memory. However, it still cannot sanitize the neighboring data (Chen et al. (2016) provides a concrete structural artifacts introduced by the deleted data. Chen attack scenario which can take advantage of the structural et al. (2016) investigated another novel security notion, artifacts to recover bitcoin transactions). namely, truly secure deletion, which can ensure the saniti- To remove impacts of the past existence of the deleted data (including both structural artifacts and side effects), zation of both the data and the structural artifacts. a few recent works investigate undetected secure dele- To achieve truly secure deletion, Chen et al. (2016)pro- tion, truly secure deletion, history independence, as well posed TedFlash, a truly secure deletion scheme for NAND as untraceable deletion. flash-based block devices. In TedFlash, the data of every write to NAND flash will be placed to an empty loca- Undetectable secure deletion. Jia et al. (2016)tried to tion which is randomly selected. Note that as the random achieve the secure deletion guarantee by hiding the dele- placement of data is independent and does not affect tion history. Intuitively, if the deletion history is concealed theplacementsofany otherdata, TedFlash canelimi- from the adversary, he/she should not be able to find nate the structural artifacts brought by each write. Most out whether there was a deletion in the past, eliminating importantly, the random placement technique is exclu- his/her possibility in recovering the deleted data or learn- sively feasible for NAND flash, because: 1) Random seeks ing anything about the deleted data. Based on this key on flash memory are as efficient as sequential seeks. 2) observation, Jia et al. investigated a novel security notion The random placements can distribute data evenly among for NAND flash-based block devices, i.e., undetectable flash, naturally achieving a good wear leveling. secure deletion, to achieve two security properties: 1) Data to be deleted are completely removed from NAND flash History independence. History independence is pro- memory, which ensures that the adversary cannot have posed to prevent historic information about the pattern of access to thedataoncetheyhavebeendeleted;2)The access to a data structure from being leaked through its deletion history is concealed from the adversary, which representation when observed by an adversary (Chen and ensures that the adversary cannot gain any knowledge Sion 2015). History independence ensures that by hav- about whether there was a deletion in the past. ing access to a storage state, the adversary is not able to To realizeundetectablesecuredeletion, Jiaetal. pro- identify the operation sequence which leads to this state. posed NAND Flash Partial Scrubbing (NFPS), an unde- In other words, given two operation sequences leading to tectable secure deletion scheme for NAND flash-based thesamestorage state: onesequencehas adeleteopera- block devices. Having observed that deleting data from tion (e.g., delete D) and its corresponding insert operation Zhang et al. Cybersecurity (2018) 1:1 Page 17 of 20 (e.g., insert D), and the other sequence does not have the Untraceable deletion. Bajaj and Sion (2013a)introduced aforementioned delete and insert operation, the adversary untraceable secure deletion, aiming to remove side effects will not be able to differentiate which operation sequence of the data being deleted (i.e., the impacts of the past led to this storage state. Therefore, history independence existence of the deleted data on the data which are still ensures that after having removed a data record, the stor- preserved in the computing devices). The corresponding age state is somehow equivalent to a state that the deleted design, Ficklebase, is to achieve untraceable secure dele- record never exists, naturally achieving the secure dele- tion in the context of relational databases. In Ficklebase, tion guarantee. Implicitly, history independence ensures once a tuple is expired, all its side-effects will be removed that no structural artifacts will be introduced. Other- via versioning and query rewriting. wise, the memory representation of each storage state Table 5 summarizes the secure deletion approaches cannot be “independent” of the operation sequence which handle impacts of the past existence of deleted leading to it. data, including both structural artifacts and side effects. Bajaj and Sion (2016; Bajaj and Sion 2013b)designed The approaches can also securely remove data themselves, history independent file system (HIFS), the first approach and may be incorporated into different layers of a storage which can provide history independence for file storage system. over mechanical hard disks. HIFS uses a history indepen- dent hash table to allocate file data to the underlying block Future directions device in a history independent manner. Most signifi- This section provides an overview of the promising cantly, they can simultaneously achieve history indepen- research directions for both PDE and secure deletion. dence and preserve locality to improve I/O performance. However, HIFS does not work for flash-based storage Achieving plausible deniability and secure deletion systems (“The architecture of a storage system”section), in a single system. Most of the existing systems either since flash memory usually has its own internal software provide deniability or achieve secure deletion. However, layer (e.g., FTL) which is used to transparently handle its data confidentiality should be simultaneously ensured special characteristics by using a special history depen- during and after the lifetime of the data, because: First, dent block placement technique. by recovering the data being deleted, the adversary can Chen et al. thus proposed HiFlash (Chen and Sion 2015), achieve a similar gain comparable to compromising con- aiming to achieve history independence in flash-based fidentiality of the data being stored; Second, if the con- block devices. To achieve history independence, HiFlash fidentiality of the data cannot be ensured during their enforces a bijection between block device and flash mem- lifetime, secure deletion (i.e., ensuring confidentiality of ory. Specifically, HiFlash always places data records, which the data after their lifetime) turns meaningless since are written to the same block-device locations, to the the adversary has already obtained the data before they same flash locations, regardless of their write patterns. are “securely” removed. Therefore, we expect a system However, by introducing a bijective mapping between which can achieve both plausible deniability and secure block device and flash memory, HiFlash can only remove deletion. Simply combining the existing PDE and secure structural artifacts introduced by the software component deletion may be problematic, since secure deletion may staying between the block device and the flash mem- require a fine-grained encryption mechanism, and plau- ory (e.g., FTL). It is unfortunately not able to remove sibly deniable encryption is not necessarily designed as the structural artifacts introduced from the upper layers. fine-grained. In addition, when pre-processing data for Therefore, it strongly relies on the assumption that the secure deletion purpose, careful consideration may be upper layer has eliminated the structural artifacts, which needed to avoid bringing in any deniability compromises. is not necessarily true. Theonlyattempt forthistypeofsystemisDEFY(Peters Table 5 Comparison of features offered by secure deletion approaches designed to handle impacts created by past existence of the deleted data NFPS (Jia et al. 2016) TedFlash (Chen HIFS (Bajaj and HiFlash (Chen and Sion 2015) Ficklebase (Bajaj and Sion 2013a) et al. 2016) Sion 2013b; Bajaj et al. 2016) Method to Overwriting Encryption Overwriting Overwriting Encryption sanitize data Method to Partial scrubbing Random History indepen- One-one mapping Versioning, query rewriting sanitize past placement dent hash table impacts technique Layer Controller (FTL) Controller (FTL) File system Controller (FTL) Application Zhang et al. Cybersecurity (2018) 1:1 Page 18 of 20 et al. 2015), but DEFY itself may suffer from deniability Authors’ contributions All authors read and approved the final manuscript. compromise (Jia et al. 2017). Competing interests Providing confidentiality guarantee for light-weight The authors declare that they have no competing interests. computing devices. Computing devices nowadays are turning more and more light-weight. Wearable devices Publisher’s Note like smart watches and smart glasses, IoT devices like Springer Nature remains neutral with regard to jurisdictional claims in smart home hubs or smart plugs, are increasingly popular published maps and institutional affiliations. today. Those light-weight computing devices are usually equipped with limited computational power. However, Author details both the PDE and the secure deletion usually require Data Assurance and Communication Security Research Center, Chinese Academy of Sciences, Beijing, China. State Key Laboratory of Information expensive encryption operations, and thus cannot directly Security, Institute of Information Engineering, Chinese Academy of Sciences, fit the use of light-weight devices. This can be mitigated 3 Beijing, China. School of Cyber Security, University of Chinese Academy of by either outsourcing part of the expensive computation Sciences, Beijing, China. School of Information Systems, Singapore Management University, Singapore, Singapore. Department of Computer to the third-party cloud providers (without confidentiality Science, Michigan Technological University, Houghton, USA. compromise) or reducing the level of security to improve performance. Received: 19 January 2018 Accepted: 17 April 2018 Eliminating deniability compromise and data leakage. The existing PDE/secure deletion systems mainly focus References on external storage, and may neglect the security leakage Amazon: New SSD-Backed Elastic Block Storage (2016). Retrieved December in other sources like memory. By having obtained snap- 17, 2017, from https://aws.amazon.com/blogs/aws/new-ssd-backed- elastic-block-storage/ shots of the victim computing device, the adversary may Anderson R, Needham R, Shamir A (1998) The steganographic file system. capture the memory state, and may compromise deni- In: International Workshop on Information Hiding. Springer. ability or recover the data being deleted by performing pp 73–82 forensic analysis. Therefore, processing of the sensitive Apple: Use filevault to encrypt the startup disk on your mac (2017). Retrieved December 17, 2017, from https://support.apple.com/en-us/HT204837 data should be conducted in an isolated memory region, Assange J, Weinmann RP, Dreyfus S (2012) Rubberhose: Cryptographically which cannot be learned by the adversary. This isolated deniable transparent disk encryption system. Retrieved December 17, memory region can be created by using trusted execution 2017, from https://web.archive.org/web/20120716034441/http:// marutukku.org/ environment (e.g., Intel SGX (Intel: Intel software guard Bajaj S, Chakraborti A, Sion R (2016) Practical foundations of history extensions 2017)). independence. Information Forensics and Security. IEEE Transactions on 11(2):303–312 Secure against quantum computing. Both PDE and Bajaj S, Sion R (2013) Ficklebase: Looking into the future to erase the past. In: Data Engineering (ICDE) 2013 IEEE 29th International Conference on. secure deletion may rely on encryption (e.g., AES-128, IEEE. pp 86–97 XTS-AES), which is not necessarily quantum resistant. Bajaj S, Sion R (2013) Hifs: History independence for file systems. The development of quantum computers seems to be In: Proceedings of the 2013 ACM SIGSAC conference on Computer & communications security. ACM. pp 1285–1296 expedited recent years (Wikipedia: Timeline of quan- Ballard Spahr LLP: State and local governments move swiftly to sue equifax tum computing 2018). Therefore, there will be a need (2017). Retrieved December 17, 2017, from https://www.natlawreview. to ensure that the PDE and the secure deletion are com/article/state-and-local-governments-move-swiftly-to-sue-equifax Bauer S, Priyantha NB (2001) Secure data deletion for linux file systems. In: secure against quantum computing, which requires care- Usenix Security Symposium. Volume 174 fully checking the existing cryptography primitives being Blass EO, Mayberry T, Noubir G, Onarlioglu K (2014) Toward robust hidden used and replacing those which are vulnerable to quantum volumes using write-only oblivious ram. In: Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security. ACM. computing with the ones that are quantum resistant. pp 203–214 Boneh D, Lipton RJ (1996) A revocable backup system. In: USENIX Security Conclusion Symposium. pp 91–96 In this survey, we summarize the existing approaches Braga A, Colito A (2014) Adding Secure Deletion to an Encrypted File System on Android Smartphones. In: SECURWARE 2014, The Eighth International for protecting data confidentiality against snapshot adver- Conference on Emerging Security Information, Systems and Technologies. saries. Our survey covers techniques for both PDE and pp 106–110 secure deletion, which can ensure data confidentiality Breeuwsma M, De Jongh M, Klaver C, Van Der Knijff R, Roeloffs M (2007) Forensic data recovery from flash memory. Small Scale Digit Device during and after their lifetime, respectively. We also dis- Forensic J 1(1):1–17 cuss a few promising future directions. Cai Y, Luo Y, Ghose S, Mutlu O (2015) Read disturb errors in mlc nand flash memory: Characterization, mitigation, and recovery. In: Dependable Systems and Networks (DSN), 2015 45th Annua IEEE/IFIP International Acknowledgment Conference on. IEEE. pp 438–449 This work was partially supported by the National Key Research & Development Canetti R, Dwork C, Naor M, Ostrovsky R (1997) Deniable encryption. In: Annual Program of China (Grant No. 2017YFC0822704) and National Natural Science International, Cryptology Conference. Springer. pp 90–104 Foundation of China (No. 61602476, No. 61772518 and No. 61602475). Zhang et al. Cybersecurity (2018) 1:1 Page 19 of 20 Cbsnews: Apple’s celebrity icloud leak probably has mundane causes (2014). Intel Corporation: Intel Solid-State Drive Optimizer (2009). Retrieved December Retrieved December 17, 2017, from https://www.cbsnews.com/news/ 17, 2017, from http://download.intel.com/design/flash/nand/mainstream/ apples-celebrity-icloud-leak-probably-has-mundane-causes/ IntelSSDOptimizerWhitePaper.pdf Chakraborti A, Chen C, Sion R (2016) Poster: Datalair: A storage block device Jagdmann D (2015) srm - Linux man page. Retrieved December 17, 2017, from with plausible deniability. ACM https://www.mankier.com/1/srm Chakraborti A, Chen C, Sion R (2017) Datalair: Efficient block storage with Jia S, Xia L, Chen B, Liu P (2016) Nfps: Adding undetectable secure deletion to plausible deniability against multi-snapshot adversaries. Proc Priv Enhanc flash translation layer. In: Proceedings of the 11th ACM on Asia Conference Technol 3:175–193 on Computer and Communications Security. ACM. pp 305–315 Chang B, Cheng Y, Chen B, Zhang F, Zhu WT, Li Y, Wang Z (2018) User-friendly Jia S, Xia L, Chen B, Liu P (2017) Deftl: Implementing plausibly deniable deniable storage for mobile devices. Comput Secur 72:163–174 encryption in flash translation layer. In: Proceedings of the 24th ACM Chang B, Wang Z, Chen B, Zhang F (2015) Mobipluto: File system friendly conference on Computer and communications security. ACM deniable storage for mobile devices. In: Proceedings of the 31st Annual Joukov N, Papaxenopoulos H, Zadok E (2006) Secure deletion myths, issues, Computer Security Applications Conference. ACM. pp 381–390 and solutions. ACM Work Storage Secur Survivability:61–66 Chang B, Zhang F, Chen B, Li Y, Zhu WT, Tian Y, Wang Z, Ching A (2018) Joukov N, Zadok E (2005) Adding secure deletion to your favorite file system. Mobiceal: Towards secure and practical plausibly deniable encryption on In: Third IEEE International Security in Storage Workshop (SISW’05). mobile devices. In: Proceedings of the 48th IEEE/IFIP International IEEE. p 8 Conference on Dependable Systems and Networks Klonowski M, Kubiak P, Kutyłowski M (2008) Practical deniable encryption. Chen B, Jia S, Xia L, Liu P (2016) Sanitizing data is not enough!: towards SOFSEM 2008: Theory Pract Comput Sci:599–609 sanitizing structural artifacts in flash media. In: Proceedings of the 32nd Lee B, Son K, Won D, Kim S (2011) Secure data deletion for usb flash memory. Annual Conference on Computer Security Applications. ACM. pp 496–507 J Inf Sci Eng 27(3):933–952 Chen B, Sion R (2015) Hiflash: A history independent flash device. arXiv Lee C, Sim D, Hwang JY, Cho S (2015) F2fs: a new file system for flash storage. preprint arXiv:1511.05180 Usenix Conf File Storage Technol:273–286 Congress US (1996) HealthInsurancePortability and Accountability Act. Retrieved Lee J, Heo J, Cho Y, Hong J, Shin SY (2008) Secure deletion for nand flash file November 9, 2017, from http://www.hhs.gov/ocr/privacy/index.html system. In: Proceedings of the 2008 ACM symposium on Applied Congress US (1999) Gramm-Leach-Bailey Act. Retrieved November 9, 2017, computing. ACM. pp 1710–1714 from http://www.gpo.gov/fdsys/pkg/PLAW-106publ102/pdf/PLAW- Lee J, Yi S, Heo J, Park H, Shin SY, Cho Y (2010) An efficient secure deletion 106publ102.pdf scheme for flash file systems. J Inf Sci Eng 26(1):27–38 Czeskis A, Hilaire DJS, Koscher K, Gribble SD, Kohno T, Schneier B (2008) LSoft Technologies Inc: Active@ KillDisk (2017). Retrieved December 17, 2017, Defeating encrypted and deniable file systems: Truecrypt v5. 1a and the from http://www.killdisk.com/ case of the tattling os and applications. In: HotSec McDonald AD, Kuhn MG (2000) Stegfs: A steganographic file system for linux. Durak B (2006) wipe - Linux man page. Retrieved December 17, 2017, from In: Information Hiding. Springer. pp 463–477 https://linux.die.net/man/1/wipe Memory Technology Devices: UBIFS (2015). Retrieved December 17, 2017, Diesburg S, Meyers C, Stanovich M, Mitchell M, Marshall J, Gould J, Wang AIA, from http://www.linux-mtd.infradead.org/doc/ubifs.html Kuenning G (2012) Trueerase: Per-file secure deletion for the storage data Meng B, Wang Jq (2009) A receiver deniable encryption scheme. In: path. In: Proceedings of the 28th, Annual Computer Security Applications International Symposium on Information Processing (ISIP ’09) Conference. ACM. pp 439–448 Meng B, Wang J (2010) An efficient receiver deniable encryption scheme and Dürmuth M, Freeman DM (2011) Deniable encryption with negligible its applications. JNW 5(6):683–690 detection probability: An interactive construction. In: Annual International Microsoft: Bitlocker (2017). Retrieved December 17, 2017, from https://docs. Conference on the Theory and Applications of Cryptographic Techniques. microsoft.com/en-us/windows/device-security/bitlocker/bitlocker- Springer. pp 610–626 overview Frèche G, Bloch MR, Barret M (2017) Polar codes for covert communications Min C, Kim K, Cho H, Lee SW, Eom YI (2012) Sfs: random write considered over asynchronous discrete memoryless channels. Entropy 20(1):3 harmful in solid state drives. In: FAST. p 12 Garfinkel SL, Shelat A (2003) Remembrance of data passed: A study of disk O’Neill A, Peikert C, Waters B (2011) Bi-deniable public-key encryption. Adv sanitization practices. IEEE Secur Priv 99(1):17–27 Cryptol–CRYPTO 2011:525–542 Geambasu R, Kohno T, Levy AA, Levy HM (2009) Vanish: Increasing data privacy Pang H, Tan KL, Zhou X (2003) Stegfs: A steganographic file system. In: Data with self-destructing data. In: USENIX Security Symposium. pp 299–316 Engineering, 2003. Proceedings. 19th International Conference on. IEEE. GEEP EDS LLC: Darik’s Boot and Nuke (2017). Retrieved December 17, 2017, pp 657–667 from http://www.dban.org/ Pang H, Tan KL, Zhou X (2004) Steganographic schemes for file system and b-tree. IEEE Trans Knowl Data Eng 16(6):701–713 Goldreich O, Ostrovsky R (1996) Software protection and simulation on oblivious rams. J ACM (JACM) 43(3):431–473 Perlman R (2005) The Ephemerizer: Making Data Disappear. In: Journal of Gutmann P (1996) Secure deletion of data from magnetic and solid-state Information System Security. Citeseer memory. In: Proceedings of the Sixth USENIX Security Symposium, San Perlman R (2005) File system design with assured delete. In: Third IEEE Jose, CA. Volume 14 International Security in Storage Workshop (SISW’05). IEEE. p 6 Halderman JA, Schoen SD, Heninger N, Clarkson W, Paul W, Calandrino JA, Peters TM, Gondree MA, Peterson ZN (2015) DEFY: A deniable, encrypted file Feldman AJ, Appelbaum J, Felten EW (2009) Lest we remember: cold-boot system for log-structured storage. In: 22th Annual Network and Distributed attacks on encryption keys. Commun ACM 52(5):91–98 System Security Symposium. NDSS Han J, Pan M, Gao D, Pang H (2010) A multi-user steganographic file system on Peterson ZN, Burns RC, Herring J, Stubblefield A, Rubin AD (2005) untrusted shared storage. In: Proceedings of the 26th Annual Computer Secure deletion for a versioning file system. In: FAST. Volume 5. Security Applications Conference. ACM. pp 317–326 pp 143–154 Howlader J, Basu S (2009) Sender-side public key deniable encryption scheme. Plumb C (2010) shred 1 - Linux man page. Retrieved December 17, 2017, from In: Advances in Recent Technologies in Communication and Computing, https://linux.die.net/man/1/shred 2009. ARTCom’09. International Conference on. IEEE. pp 9–13 Qin Y, Tong W, Liu J, Zhu Z (2013) Smsd: A smart secure deletion scheme for Hu J, Yan S, Zhou X, Shu F, Wang J (2017) Covert Communication in Wireless ssds. J Converg 4:30–35 Relay Networks. In: GLOBECOM 2017 - 2017 IEEE Global Communications Kissel R, Scholl M, S S, Li X (2006) Guidelines for Media Sanitization. Retrieved Conference. pp 1–6 December 17, 2017, from http://ws680.nist.gov/publication/get_pdf.cfm? Ibrahim MH (2009) Receiver-deniable public-key encryption. IJ Netw Secur pub_id=50819 8(2):159–165 Reardon J, Basin D, Capkun S (2013) Sok: Secure data deletion. In: Security and Incits: Scsi storage interfaces (2016). Retrieved December 17, 2017, from http:// Privacy (SP), 2013 IEEE Symposium on. IEEE. pp 301–315 www.t10.org/index.html Reardon J, Capkun S, Basin D (2012) Data node encrypted file system: Efficient Intel: Intel software guard extensions (2017). Retrieved December 19, 2017, secure deletion for flash memory. In: Proceedings of the 21st USENIX from https://software.intel.com/en-us/sgx conference on Security symposium, USENIX Association. pp 17–17 Zhang et al. Cybersecurity (2018) 1:1 Page 20 of 20 Reardon J, Marforio C, Capkun S, Basin D (2012) Secure deletion on Zhou X, Pang H, Tan KL (2004) Hiding data accesses in steganographic file log-structured file systems. ASIACCS 2012 system. In: Data Engineering, 2004. Proceedings. 20th International Reardon J, Ritzdorf H, Basin D, Capkun S (2013) Secure data deletion from Conference on. IEEE. pp 572–583 persistent media. In: ACM Sigsac Conference on, Computer & Zuck A, Shriki U, Porter DE, Tsafrir D (2017) Preserving hidden data with an Communications Security. pp 271–284 ever-changing disk. Work Hot Top Oper Syst Robust Flash Storage: YAFFS (2002). Retrieved December 17, 2017, from http:// www.yaffs.net/yaffs-overview Roche DS, Aviv A, Choi SG, Mayberry T (2017) Deterministic, stash-free write-only oram. In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security. ACM. pp 507–521 Sarbanes USP, Oxley URMG (2002) Sarbanes-Oxley Act. Retrieved November 9, 2017, from http://www.sec.gov/about/laws.shtml#sox2002 Skillen A. (2013) Deniable storage encryption for mobile devices. PhD thesis. Concordia University Skillen, A, Mannan M (2013) On implementing deniable storage encryption for mobile devices Skillen A, Mannan M (2014) Mobiflage: Deniable storage encryptionfor mobile devices. IEEE Trans Dependable Secure Comput 11(3):224–237 Sourceforge: FreeOTFE A free “on-the-fly” transparent disk encryption program for PC and PDAs (2017). Retrieved December 19, 2017, Project website: https://sourceforge.net/projects/freeotfe.mirror/ Sourceware: Jffs2 (2003). Retrieved November 9, 2017, from https://www. sourceware.org/jffs2/ SQLite: Pragma statements (2017). Retrieved December 17, 2017, from http:// www.sqlite.org/pragma.html#pragma_secure_delete Stahlberg P, Miklau G, Levine BN (2007) Threats to privacy in the forensic analysis of database systems. ACM SIGMOD conf Manag data:91–102 Stefanov E, Van Dijk M, Shi E, Fletcher C, Ren L, Yu X, Devadas S (2013) Path oram: an extremely simple oblivious ram protocol. In: Proceedings of the 2013 ACM SIGSAC conference on Computer & communications security. ACM. pp 299–310 Sun K, Choi J, Lee D, Noh SH (2008) Models and design of an adaptive hybrid scheme for secure deletion of data in consumer electronics. Consumer Electronics. IEEE Trans 54(1):100–104 Swanson S, Wei M (2010) Safe: Fast, verifiable sanitization for ssds. University of California-San Diego, San Diego, CA Tang Y, Lee PP, Lui J, Perlman R (2012) Secure overlay cloud storage with access control and assured deletion. Dependable and Secure Computing. IEEE Trans 9(6):903–916 Team work systems: Advanced technology attachment (2017). Retrieved December 17, 2017, from https://teamws.com/glossary/ata-advanced- technology-attachment/ Thornber J Thin provisioning tools. https://github.com/jthornber/thin- provisioning-tools Troncoso C, Diaz C, Dunkelman O, Preneel B (2007) Traffic analysis attacks on a continuously-observable steganographic file system. In: International Workshop on Information Hiding. Springer. pp 220–236 TrueCrypt: Free open source on-the-fly disk encryption software.version 7.1a (2012). Retrieved December 19, 2017, Project website: http://www. truecrypt.org/ Varun S, Shibin K, Anoop S, Vivek K (2017) Magikfs - The Steganographic Filesystem On Linux. Retrieved December 19, 2017, Project website: http:// magikfs.sourceforge.net/ Wang Z, Murmuria R, Stavrou A (2012) Implementing and Optimizing an Encryption Filesystem on Android. In: IEEE International Conference on Mobile Data Management. pp 52–62 Wei MYC, Grupp LM, Spada FE, Swanson S (2011) Reliably erasing data from flash-based solid state drives. In: FAST. Volume 11. pp 8–8 Wikipedia: Timeline of quantum computing (2018). Retrieved December 19, 2017, from https://en.wikipedia.org/wiki/Timeline_of_quantum_ computing Yu X, Chen B, Wang Z, Chang B, Zhu WT, Jing J (2014) Mobihydra: Pragmatic and multi-level plausibly deniable encryption storage for mobile devices. In: Information Security - 17th International Conference, ISC 2014, Hong Kong, China, October 12-14 2014. Proceedings. pp 555–567 Zarras A, Kohls K, Dürmuth M, Pöpper C (2016) Neuralyzer: Flexible expiration times for the revocation of online data. In: Proceedings of the Sixth ACM on Conference on Data and Application Security and Privacy. ACM. pp 14–25 Zhao L, Mannan M (2015) Gracewipe: Secure and verifiable deletion under coercion. In: Network and Distributed System Security Symposium http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Cybersecurity Springer Journals

Ensuring data confidentiality via plausibly deniable encryption and secure deletion – a survey

Free
20 pages

Loading next page...
 
/lp/springer_journal/ensuring-data-confidentiality-via-plausibly-deniable-encryption-and-EqE815hkre
Publisher
Springer Singapore
Copyright
Copyright © 2018 by The Author(s)
Subject
Computer Science; Computer Science, general
eISSN
2523-3246
D.O.I.
10.1186/s42400-018-0005-8
Publisher site
See Article on Publisher Site

Abstract

Ensuring confidentiality of sensitive data is of paramount importance, since data leakage may not only endanger data owners’ privacy, but also ruin reputation of businesses as well as violate various regulations like HIPPA and Sarbanes-Oxley Act. To provide confidentiality guarantee, the data should be protected when they are preserved in the personal computing devices (i.e., confidentiality during their lifetime); and also, they should be rendered irrecoverable after they are removed from the devices (i.e., confidentiality after their lifetime). Encryption and secure deletion are used to ensure data confidentiality during and after their lifetime, respectively. This work aims to perform a thorough literature review on the techniques being used to protect confidentiality of the data in personal computing devices, including both encryption and secure deletion. Especially for encryption, we mainly focus on the novel plausibly deniable encryption (PDE), which can ensure data confidentiality against both a coercive (i.e., the attacker can coerce the data owner for the decryption key) and a non-coercive attacker. Keywords: Data confidentiality, Plausibly deniable encryption, Secure deletion Introduction Spahr LLP: State and local governments move swiftly to Modern computing devices (e.g., desktops, laptops, smart sue equifax 2017); Third, it will directly violate regulations phones, tablets, wearable devices) are increasingly used like HIPAA (Congress 1996), Gramm-Leach-Bliley Act to process sensitive or even mission critical data. Protect- (Congress 1999), and Sarbanes-Oxley Act (Sarbanes and ing confidentiality of those sensitive data is of paramount Oxley 2002). The data confidentiality should be ensured importance because: First, data leakage will endanger data not only during their lifetime (i.e., the data are preserved owners’ privacy. For example, the data leakage of iCloud in the devices), but also after their lifetime (i.e., the data in 2014 disclosed almost 500 private pictures of various have been removed from the devices). This is because, celebrities (Cbsnews: Apple’s celebrity icloud leak proba- by recovering sensitive data which have been deleted, the bly has mundane causes 2014). Second,itwillruinrepu- attacker can achieve a similar gain compared to success- tation of businesses. For example, Equifax data breach in fully attacking the confidentiality of the data being pre- July 2017 caused a leak of 145,500,000 consumer records; a served in the devices. For example, by recovering a naked few local governments like cities of Chicago and San Fran- picture deleted by a victim, the adversary can still use it to cisco, as well as the Commonwealth of Massachusetts, embarrass the victim or ask the victim for ransom money. have filed enforcement actions against Equifax (Ballard Correspondingly, the research efforts for protecting data confidentiality can be divided into two categories: encryption and secure deletion. Encryption can protect *Correspondence: jiashijie@iie.ac.cn; bchen@mtu.edu Data Assurance and Communication Security Research Center, Chinese confidentiality of the data stored at rest by transform- Academy of Sciences, Beijing, China ing them into another format using some secrets (e.g., State Key Laboratory of Information Security, Institute of Information keys), such that the adversary is not able to correlate Engineering, Chinese Academy of Sciences, Beijing, China the transformed format to the original format without Department of Computer Science, Michigan Technological University, Houghton, USA obtaining the secrets. All types of existing encryption Full list of author information is available at the end of the article mechanisms like symmetric encryption and asymmetric © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Zhang et al. Cybersecurity (2018) 1:1 Page 2 of 20 encryption can achieve the aforementioned security prop- against both the coercive adversaries and the non-coercive erty. Secure deletion is to ensure that once the sensitive adversaries. We summarize the research for PDE the- data are deleted, the probability of recovering them is ory and systems (including both the desktop systems and negligibly small. This requires special techniques to com- the mobile systems). For the second case, we summa- pletely destroy data, eliminating any traces which may lead rize secure deletion approaches in various storage media to a full/partial data recovery. includingharddiskdrives(HDDs)and NAND flashmem- To protect confidentiality of the data stored in a com- ory. We also outline the new direction of secure deletion puting device, conventional encryption may not work approaches by eliminating impacts of operation history. when both the computing device and the device’s owner are captured by an attacker, since the attacker can coerce Organization. In “Background”section,weintroduce the owner to disclose the secret (i.e., a coercive adver- the background knowledge of flash memory, the architec- sary). Once the secret is disclosed, the transformed format ture of a storage system, PDE as well as secure deletion. created by encryption will be reversed, and the sensitive In “Models and assumptions” section, we unify the adver- data will be leaked. A novel encryption technique, plau- sarial model for both PDE and secure deletion, and also sibly deniable encryption (PDE) (Canetti et al. 1997), has summarize the assumptions required by PDE. We then been designed to complement the traditional encryption summarize the literature for PDE in “Protecting data con- to handle such a coercive attack. The high-level idea of fidentiality against coercive adversaries via PDE”section PDE is, the original sensitive message is encrypted into and for secure deletion in “Ensuring confidentiality of ciphertexts in a special way, such that during decryption, the deleted data via secure deletion”section,respectively. if a true key is used, the original sensitive message can In “Future directions”section,wediscuss afew future be recovered, but if a decoy key is used, a plausible mes- directions of PDE and secure deletion. We conclude in sage will be generated. Therefore, upon being coerced, “Conclusion”section. the owner can simply disclose the decoy key to protect confidentiality of the original sensitive message. Background To protect confidentiality of the data deleted from Flash memory a computing device, the deleted data should be made Flash memory is a solid-state non-volatile computer stor- completely unrecoverable. Conventionally, this is ensured age medium that can be electrically erased and repro- by carefully over-writing the storage medium storing grammed. It can eliminate mechanical limitations and the data using garbage information (Joukov and Zadok latency of hard drives, achieving a much higher I/O 2005; Wei et al. 2011; Garfinkel and Shelat 2003;Sun throughput with a much lower power consumption. et al. 2008; Gutmann 1996) or deploying encryption Therefore, it gains popularity in main-stream mobile using ephemeral keys (Perlman 2005a, b,Geambasu devices like smart phones, tablets, wearable devices. In et al. 2009,Tangetal. 2012, Reardon et al. 2012,Zarras addition, a lot of high-end laptops like Apple MacBook use et al. 2016). This unfortunately was shown to be insuf- flash memory as external storage. Even cloud providers ficient, since past existence of the deleted data will cre- allow their users to choose solid state drives (SSDs) as ate impacts on both the data organization (Bajaj and the underlying storage media (Amazon: New SSD-Backed Sion 2013b) and the other data which have not been Elastic Block Storage 2016). The flash memory being used deleted (Bajaj and Sion 2013a). Those impacts can then as the storage medium is mainly NAND flash. NAND be utilized by the adversary as an oracle to derive sen- flash stores data using an array of memory cells, which are sitive information about the deleted data. In the worst grouped into pages (each page can store 512-byte, 2KB, case, the adversary is able to completely recover the data or 4KB data), and multiple pages are further grouped into being deleted (Chen et al. 2016). Therefore, recent secure blocks (can contain 32, 64, or 128 pages). deletion approaches focus on eliminating those impacts Compared to traditional mechanical disks, NAND flash (Bajaj and Sion 2013a, b,ChenandSion 2015;Chenand has several unique characteristics. First, NAND flash has Sion 2016;Jia et al. 2016). an erase-before-write design. Specifically, to overwrite In this work, we aim to conduct a thorough literature a flash page, the page needs to be erased before any review on data confidentiality protection. We believe that new data can be programmed to it. Second, the unit of data confidentiality should be always ensured no mat- read/program of NAND flash is a page, but the unit of terthedata arepreserved in thecomputing device or erase is a block, which consists of multiple pages. The have been removed. Therefore, our survey covers tech- first and the second special characteristics of NAND flash niques being used to protect data confidentiality for both make it expensive to perform an in-place update. There- cases: 1) data are stored in the devices; 2) data have been fore, flash memory usually prefers an out-of-place update. removed from the devices. For the first case, we mainly Third, each flash block has a finite number of program- focus on PDE which can provide confidentiality guarantee erase (P/E) cycles. In other words, a flash block will be Zhang et al. Cybersecurity (2018) 1:1 Page 3 of 20 worn out if the number of programs/erasures performed over it exceeds a certain threshold. Last, NAND flash is vulnerable to read/program disturb (Cai et al. 2015). In other words, frequently reading/writing the same flash location may corrupt the data stored nearby. Due to read/program disturb (Cai et al. 2015), NAND flash usu- ally requires computing an error-correcting code (ECC) for each page, and storing the parity in the “spare area” of the corresponding page. In general, the raw flash can be managed through either flash translation layer (FTL) or flash-specific file system. – Flash translation layer. To be compatible with traditional block-based file systems (e.g., EXT4), flash memory can be emulated as a block device by exposing a block-based access interface, which is the most popular form of flash-based products (e.g., SSDs, eMMCs, USB sticks). This is usually achieved Fig. 1 The architecture of storage systems by introducing special firmware, flash translation layer (FTL), between the file system and the raw flash. FTL can translate the logical block addresses to the underlying physical flash addresses, providing a block The lowest layer is the physical medium layer, where access interface to upper layers. data are actually stored, e.g., HDDs or NAND flash. The – Flash file systems. Another alternative of using raw physical storage medium is always accessed through a flash is to directly build a flash-specific file system controller. The basic function of the controller is to trans- over it. A flash file system is a file system optimized late data format on the physical storage medium (e.g., specifically for flash memory. Popular flash file electrical voltage) into another format (e.g., binary values) systems include YAFFS (Robust Flash Storage: understandable by upper layers. The controller offers a YAFFS 2002), UBIFS (Memory Technology Devices: standardized and well-defined hardware interface, e.g., UBIFS 2015), JFFS2 (Sourceware: Jffs2 2003), and ATA (Team work systems: Advanced technology attach- F2FS (Lee et al. 2015). Note that flash file systems ment 2017) and SCSI (Incits: Scsi storage interfaces 2016), become less popular nowadays. Most of the recent which allows data to be read from/written to the physical mobile devices are only designed to be compatible storage medium. The HDD adopts in-place updates, and with FTL-based flash devices, and usually do not hence its controller usually consistently maps a logical allow directly accessing the raw flash. For example, block address to a certain storage location on the physical the Google Nexus 6P Android phone uses eMMC storage medium. On the contrary, NAND flash prefers cards as storage media, and only the old Android out-of-place updates due to its special features, and is usu- phones like Nexus One and Nexus S allow directly ally managed through FTL or a flash specific file system accessing the raw flash. (“Flash memory” section). Device drivers are to consol- idate access to different types of hardware by exposing Special internal management (being incorporating into a common simple interface in the form of software. The FTL or flash file systems) is usually required to han- block device driver interface allows reading and writing dle the characteristics of flash memory, which may easily of blocks in logical addresses. The block device driver can lead to deniability compromise or data leakage, making beused on topofanHDD controlorNANDflash being it challenging to provide deniability/secure deletion on encapsulated by FTL. The memory technology device flash-based storage systems. (MTD), another type of device driver, is used to access raw NAND flash memory directly. MTD permits reading The architecture of a storage system and writing, but blocks must be erased before being The storage media like magnetic HDDs and NAND flash written, which occurs at a large granularity. Unsorted are usually managed through a storage system, which is block images (UBI) is another interface for accessing organized into a few layers. The layers interact with each flash memory, which builds on top of MTD interface and other, and provide users a unified interface to access the simplifies some aspects of using raw flash memory. data being stored in the storage media. Figure 1 shows a File systems are responsible for organizing logical typical architecture for storage systems. sequences of data among the available blocks on the Zhang et al. Cybersecurity (2018) 1:1 Page 4 of 20 physical storage medium through the interface provided to the application layer), the system may simply mod- by the device driver. These include: 1) block file systems ify the metadata in the file system layer (e.g., changing built on top of block device, e.g., FAT32, EXT4, and NTFS; the block allocation table and invalidating the data being 2) flash file systems built on top of MTD device, e.g., deleted) to make the data appear to have been removed YAFFS (Robust Flash Storage: YAFFS 2002); 3) UBI file in the application layer. However, the actual content is system (Memory Technology Devices: UBIFS 2015) built still preserved in the physical storage medium layer and on top of UBI device. may be recovered by the adversary through disk foren- The highest layer is the application layer, offering an sics (Breeuwsma et al. 2007; Garfinkel and Shelat 2003). interface to the users. Therefore, secure deletion requires ensuring that the con- tent being deleted should become inaccessible at each Plausibly deniable encryption (PDE) layerofthe storagesystem. Conventional encryption is broadly used by individuals In addition, the past existence of the deleted data may and businesses to protect sensitive data. Major operat- leave artifacts in the data organization (Bajaj and Sion ing systems now increasingly support the use of full disk 2013b) or side effects on the other data (Bajaj and Sion encryption. For example, FileVault (Apple: Use filevault to 2013a). After the data have been deleted, those artifacts or encrypt the startup disk on your mac 2017)inMac OS side effects may be utilized by the adversary to learn sen- X 10.3 and later, BitLocker (Microsoft: Bitlocker 2017)in sitive information of the deleted data (Chen et al. 2016). Windows Vista and later, full-disk encryption in Android This also creates a barrier towards completely removing 4.4 and later, etc. However, traditional encryption may sensitive information. not work when data owners are captured by the adver- sary and coerced into disclosing their decryption keys. To Models and assumptions protect sensitive data against this type of coercive adver- A unified adversarial model for PDE and secure deletion saries, plausibly deniable encryption (Canetti et al. 1997; We unify the adversarial model for both PDE and secure Dürmuth and Freeman 2011;Howladerand Basu 2009; deletion by considering a snapshot adversary who can Ibrahim 2009;Klonowski et al. 2008; Meng and Wang have access to the state of a victim device. The adversary 2009; 2010; O’Neill et al. 2011) (PDE) can be utilized to is assumed to be not able to control the code of the victim hide the sensitive data by denying their very existence. Dif- system, i.e., no malicious code can be injected and hence ferent from conventional encryption, PDE encrypts origi- thecodeofthe victim system is secure.Eachaccess will nal sensitive message into ciphertexts in such a way that, allow the adversary to obtain a full snapshot of both the upon decryption, if a true key is used, the original sensi- external storage and the memory. The adversary is com- tive message can be recovered, but if a decoy key is used, a putationally bounded, and tries to illegitimately derive plausible message will be generated. When being coerced, sensitive information from the snapshots being captured. the victim can simply disclose the decoy key. Using the We consider that the adversary can have access to the decoy key, the adversary is able to decrypt the cipher- victim device once (i.e., a single-snapshot adversary)and text into the plausible message which is non-sensitive, multiple times (i.e., a multi-snapshot adversary). and is hence convinced that no sensitive information The single-snapshot adversary captures a lot of real- is stored. world scenarios. For example, an attacker steals a smart phone (Yu et al. 2014;Chang et al. 2015)oralaptop,or Secure deletion breaks into a data center obtaining a snapshot of a vic- Secure deletion is a technique designed to ensure com- tim server. The multi-snapshot adversary also captures a plete elimination of sensitive data once they become obso- lot of real-world scenarios. For example, an attacker peri- lete. It requires a guarantee that an adversary should odically breaks into a hotel room, obtaining a “memory neither recover the deleted data, nor learn anything about dump” of a victim’s smart phone; a border checker period- them. However, achieving such a guarantee in modern ically obtains snapshots from a victim’s smart device (Blass computing systems is a challenging task due to the com- et al. 2014;Petersetal. 2015). plication of the storage systems (“The architecture of a The only unique attack behavior for PDE is the adver- storage system” section). A modern storage system usually sary may coerce the data owner for the decryption keys. consists of multiple layers, and performing secure dele- This behavior is not applicable to secure deletion, because tion in one layer is usually not able to eliminate the data, no key for the deleted data will be preserved after the data since data leakage may be observed in other layers. For have been securely removed. example, given a Microsoft Word document, removing data from the document itself cannot guarantee that the The assumptions required by PDE deleted data really become inaccessible. Upon receiving a Since PDE systems usually require a few common assump- delete request issued by Microsoft Word (which belongs tions. We summarize these assumptions in the following. Zhang et al. Cybersecurity (2018) 1:1 Page 5 of 20 – The PDE software should be merged into the code layer cannot ensure that traces of the sensitive data stream of the device (e.g., part of the Android will not be observed by the snapshot adversary in the framework), such that its availability is widespread, underlying file system and physical storage medium layer, and an attacker cannot simply compromise especially when the sensitive data need to be updated deniability based on the availability of software over time and the adversary can obtain multiple snap- support. In addition, PDE requires changing a few shots (“A unified adversarial model for PDE and secure system components (e.g., booting process). The deletion” section). Therefore, when being applied to stor- adversary who can perform reverse engineering or age systems, rather than simply use encryption, two types dynamic analysis over those components will of PDE techniques, steganography and hidden volumes, unavoidably compromise deniability. The PDE are used to provide deniability. systems cannot defend against this type of adversary. The first type of PDE technique is steganography – The adversary will know the design of PDE. However, (Andersonetal. 1998). The basic idea of steganography is he/she does not have any knowledge on secret to hide sensitive data within regular file data. For exam- information (e.g., keys and passwords) being required ple, the sensitive data can be computed by performing to open/operate the PDE mode. an XOR operation over a few cover files (Anderson et al. – The adversary is rational and will stop coercing the 1998). A main concern of the steganography technique is device’s owner once he/she is convinced that the to avoid over-writing the hidden sensitive data, since they decryption keys have been revealed. are actually part of the regular data. This can be mitigated – The adversary can not capture a device working in by creating and storing (secretly) multiple copies of the the PDE mode or after a crash of the PDE mode. sensitive data, which in return will lead to inefficient use Otherwise, he/she can trivially retrieve sensitive of disk space. hidden data, compromising deniability. The other type of PDE technique is hidden volumes – The operating system, bootloader, baseband OS, (TrueCrypt: Free open source on-the-fly disk encryption firmware are all malware-free. In addition, the PDE software.version 7.1a 2012). The hidden volumes tech- mode is malware-free. nique works as follows: There are two encrypted volumes on the disk, a public volume and a hidden volume. The Protecting data confidentiality against coercive public volume is encrypted using a decoy key and the hid- adversaries via PDE den volume is encrypted using a hidden key (i.e., the true Ensuring confidentiality of the data being preserved in key). The public volume is placed on the entire disk and the hidden volume is usually placed from a secret offset personal computing devices can be achieved by encryp- towards the end of the disk (i.e., the hidden volume is part tion. However, traditional encryption cannot defend against coercive adversaries (“A unified adversarial model of the public volume, see Fig. 2). Note that initially the for PDE and secure deletion” section). Therefore, we entire disk is filled with random data and the data written mainly focus on plausibly deniable encryption (PDE), to the public volume should be placed sequentially from which can protect confidentiality of the data present in the beginning of the disk to reduce probability of over- the computing devices against both coercive and non- writing the hidden volume. When the victim is coerced coercive adversaries. into revealing the encryption key, he/she can disclose the decoy key, and the attacker will use the decoy key to PDE – from Theory to Practice decrypt the entire disk and cannot distinguish the hidden An ideal PDE would be a special encryption, which can volume from the random noise being filled initially, and is encrypt sensitive plaintexts into ciphertexts, such that the thus convinced that no sensitive data are stored. The hid- ciphertexts can be decrypted into either original sensitive den volumes technique can be viewed as a special type of plaintexts (using true key) or plausible non-sensitive steganography technique, which always hides the sensitive plaintexts (using decoy key). This is to ensure that one data in a contiguous region being placed at the end of the key can be disclosed when the data owner is coerced. disk and remains undetected by the coercive adversaries. However, such an ideal encryption is impractical for stor- Leveraging hidden volumes and steganography, vari- age systems because: First, the existing instantiation for ous PDE systems have been built to enable deniability on PDE results in a growing size of ciphertexts (Canetti et al. computing devices. All those works can be divided into 1997), which itself could be an indication of the existence two categories, PDE for desktop computers and PDE for of deniability. Second, a modern storage system is usually mobile devices. complicate, consisting of multiple layers, e.g., applica- tion layer, file system layer, physical storage medium PDE for desktop computers The existing PDE systems for desktop computers mainly layer (“The architecture of a storage system”section). rely on steganography and hidden volumes. In a few Simply encrypting the data using PDE in the application Zhang et al. Cybersecurity (2018) 1:1 Page 6 of 20 Fig. 2 The hidden volumes technique recent works, researchers also incorporate Oblivious same file system, data leakage may occur. A user of StegFS RAM (ORAM) in order to defend against multi-snapshot cannot deny the existence of hidden files, due to exis- adversaries. tence of the modified EXT2 driver and the external block table. However, the user can keep the number of security PDE using steganography. Steganographic file systems levels secret. have been initially proposed by Anderson et al. (1998), To further improve efficiency and reliability, Pang et al. with two alternative schemes. In their first scheme, the designed another steganographic file system (Pang et al. system has a number of cover files originally consisting of 2003; 2004). They use a bitmap to mark the blocks being random bits and the user’s files can be computed using used by the hidden files, and thus eliminate the need of a subset of cover files. This scheme uses a password and storing multiple copies of a file, alleviating the reliabil- the file name to determine which cover files are used. Its ity issues and I/O inefficiencies. However, the adversary drawback is requiring storing a large number of cover files. may be able to identify existence of hidden files, because Moreover, to accommodate files of arbitrary length, the the hidden blocks, being marked as used, do not pos- cover files must be relatively large. Their second scheme is sess a directory record. Three approaches are used to mitigate the aforementioned compromise. First, a few based on the computational security of block ciphers. The blocks, which do not store hidden data, are abandoned system is initialized by filling the entire hard disk with ran- dom data. A sensitive file is encrypted and stored at disk and marked as being used during initialization. Second, location being derived from the file name and the pass- when the system creates a new file, several additional word, and the encryption key is also derived in a similar blocks are allocated and filled with random noise. Third, manner. In this way, the adversary is not be able to dis- to prevent adversaries from identifying whether a block tinguish blocks containing hidden data from free blocks stores hidden data, the system maintains a few dummy being filled with random noise. However, when the disk hidden files and periodically updates them in the back- is increasingly filled by hidden files, collision of disk loca- ground. These mitigation approaches however, increase tions may increase, leading to significant over-writes. This overhead of disk space usage. is mitigated by writing each block to several disk locations. Zhou et al. further improved Pang et al.’s work by adding Inspired by the second scheme of Anderson et al. (1998), dummy transactions to obfuscate hidden files in cloud McDonald et al. proposed StegFS (McDonald and Kuhn storage (Zhou et al. 2004). Although the reliability and 2000), an EXT2-based file system which uses an exter- I/O efficiency have been improved, disk space overhead nal block allocation table to record entries for disk blocks. remains large. Troncoso et al. (2007) presented traffic StegFS supports a few security levels, each with a sep- analysis attacks on the file update algorithm proposed by arate password. To prevent overwriting the data from Zhou et al. (2004). Their attacks can detect file updates a security level which is closed, StegFS writes data in and reveal existence as well as location of files. Specif- a redundant manner. When multiple security levels are ically, they can detect files which occupy multiple data open, since hidden and regular files are present in the blocks with only two updates. Moreover, their attacks Zhang et al. Cybersecurity (2018) 1:1 Page 7 of 20 can also reveal files which occupy only one data block applications. Modern operating systems are not designed after a sufficient number of access operations. Han et al. to preserve deniability and may perform many unexpected proposed a dummy-relocatable steganographic (DRSteg) behaviors. As a result, even when the operating system file system (Han et al. 2010) to provide deniability in a runs properly, information relevant to hidden data may be multi-user environment. By sharing dummy data among leaked. For example, some operating systems (e.g., Win- multiple users in the system, DRSteg is able to increase dows) automatically create shortcuts to hidden files when the level of deniability being provided to individual users. they are used, and those shortcuts may be stored in reg- In DRSteg, in order to free dummy data without destroy- ular non-hidden directories. The shortcuts may contain ing user data, a novel dummy relocation mechanism is information about the hidden file, e.g., file name, loca- used to allow individual users to distinguish dummy data tion, length, access time, creation time and even volume from other users’ data. It can also prevent adversaries from serial number of the file system on which the hidden file distinguishing dummy and user data even after obtain- is stored. If the adversary localizes those shortcuts, it may ing multiple snapshots. There are also a few StegFS-based suspect existence of the hidden data, compromising deni- implementations including RubberhoseFS (Assange et al. ability. Another possible leakage source is the primary 2012) and Magikfs (Varun et al. 2017), which are no longer application. The primary applications are not necessarily maintained and the existing implementations may not be designed to preserve deniability, and may leak information compatible with modern Linux operating system. about the hidden data. For example, primary applica- tions may create redundant files to prevent data loss. If PDE using hidden volumes. Disk encryption tools like those files are not properly deleted, the content of the TrueCrypt (TrueCrypt: Free open source on-the-fly disk hidden data may be leaked. Finally, non-primary applica- encryption software.version 7.1a 2012)and FreeOTFE tions, such as desktop search applications, may access the (Sourceforge: FreeOTFE A free “on-the-fly” transparent files being stored in the hidden volume. Those applica- disk encryption program for PC and PDAs 2017) use hid- tions may cache snapshots of the hidden files and store den volumes to provide plausible deniability. TrueCrypt them for a later use. The adversary may also compromise supports user data encryption with several ciphers includ- deniability from those cached data. ing AES, TwoFish, Serpent, and a cascade of these ciphers in the XTS mode. The header of each TrueCrypt volume Other PDE systems for desktop computers. If the is filled with random data (e.g., salt) or encrypted with the adversary can capture multiple snapshots of hidden vol- header key which is derived from the corresponding pass- ume at different points of time, they can detect exis- tence of the hidden volume, by simply comparing dif- word using PBKDF2. Therefore, the entire volume appears ferent snapshots and identifying whether “free” blocks as being filled with randomness. When TrueCrypt loads have been changed. Therefore, TrueCrypt cannot pro- a volume, since it does not store the cipher specifica- vide deniability when facing a multi-snapshot adversary. tion, all supported ciphers will be tried with a header key Blass et al. proposed HIVE (Blass et al. 2014) to allow the (being derived from the user’s password using PBKDF2) user to deny existence of the hidden volume when fac- until it decrypts the volume and obtains the ASCII string ing a multi-snapshot attack. This is achieved by hiding “TRUE” from a certain block in the header. Then True- Crypt decrypts the encrypted master volume key using every access of the disk using Oblivious RAM (ORAM) the header key. Note that the master volume key is gener- (Goldreich and Ostrovsky 1996;Stefanovetal. 2013), ated randomly upon creating the volume. If a TrueCrypt which was originally designed to hide patterns of access to hidden volume is created, there will be also a hidden remote storage. However, ORAM is known as expensive header, which contains offset of the hidden volume. The in terms of both computation and I/O. Although HIVE hidden header is tested before the public header when uses a less expensive write-only ORAM that only supports mounting a volume. TrueCrypt can also create a hidden write operations, its performance is still far from being OS in a hidden volume by creating a new partition and practical. copying the current OS to the hidden volume. When the To improve performance of HIVE, Chakraborti et al. system is booted into the hidden OS, all unencrypted vol- proposed DataLair (Chakraborti et al. 2016; 2017). Hav- umes and non-hidden encrypted volumes are mounted ing observed that revealing access patterns to the public in a read-only manner, ensuring that any OS/application- data is unnecessary, DataLair only protects operations specific leakage stays within the hidden volume. on the hidden data and ensures that they are indistin- Czeskis et al. (2008) analyzed TrueCrypt and proposed guishable from operations on the public data. In addition, three types of attacks against it. They consider three DataLair optimizes the oblivious access mechanism being types of leakage sources: 1) the operating system; and deployed for hidden data. Compared to HIVE, DataLair is twoordersofmagnitude faster in termsofpublicdata 2) the primary applications (i.e., an application that is access, and five times faster in terms of hidden data access. used to manage hidden data); and 3) the non-primary Zhang et al. Cybersecurity (2018) 1:1 Page 8 of 20 Roche et al. designed DetWoORAM (Roche et al. 2017), derives the hidden volume key; and 2) decoy password an entirely new technique for write-only ORAM, which that derives only the decoy volume key; and 3) deletion uses a deterministic and sequential writing pattern, elim- password that derives the decoy volume key and over- inating the need of any “stashing” of blocks in local state. writes the hidden volume key. When coerced, the victim In DetWoORAM, since the write will always succeed and can fake compliance, and enter the deletion password, and occur in a free block, the notion of stash can be com- then can prove to the attacker that Gracewipe has been pletely removed. They also pointed out that the construc- executed and the real key is no longer available. tion of DataLair does not satisfy write-only obliviousness, Table 1 summarizes the existing PDE systems for desk- since the process of finding free blocks leaks information top computers, being incorporated into different lay- about which blocks are free, and the adversary can tell ers of the storage systems. To defend against snapshot whether recent writes have been performed on the same adversaries, they may rely on hidden volumes, steganog- address or not. raphy, or ORAM, and provide one or multiple denia- Zuck et al. presented Ever-Changing Disk (ECD) (Zuck bility levels with different overheads. Schemes based on et al. 2017) to achieve deniable storage system. Their steganography (e.g., McDonald and Kuhn 2000;Pang design follows three requirements: 1) resistance to multi- et al. 2003) suffer from high performance overhead since snapshot adversaries; and 2) ensuring that hidden data will they write multiple copies of files or dummy writes not be destroyed when a user is writing the public volume; for steganography. Hidden volume mechanism (True- and 3) using normal system operations on public data to Crypt: Free open source on-the-fly disk encryption soft- disguise writes to hidden data. In ECD, the storage space is ware.version 7.1a 2012) introduces low overhead since the separated into two parts: a part containing the public data hidden files are stored in the free space and the over- volume and the other containing the hidden data volume. write problem is mitigated by linear space allocation. The hidden volume is visible to the system when the user Since ORAM is known to be expensive, schemes based enters the secret key. ECD uses a large volume of pseu- on ORAM (e.g., Blass et al. 2014; Chakraborti et al. 2017; dorandom data to hide the sensitive data. A portion of Roche et al. 2017) unavoidably have high overheads. data from the volume are periodically migrated using nor- mal firmware operations to obfuscate writes to the hidden PDE for mobile devices data. Since hidden and pseudorandom data blocks are Compared to desktop computers, mobile devices are usu- constantly relocated and modified, the hidden data may ally different in two aspects: First, they are equipped with eventually be overwritten without knowing the secret key. less computational power. Second, they usually use flash To mitigate the overwrite issue, the rate of internal data memory (“Flash memory”section)asstorage.The exist- migration is controlled by the user and the user should ing PDE systems for mobile devices can be divided into enter the secret key periodically. two categories: 1) The PDE systems built on top of block Zhao et al. proposed Gracewipe (Zhao and Mannan devices. This type of PDE systems views flash memory as a 2015), by which the victim can provably destroy/erase data black box, which exposes a block-access interface through when being coerced, hoping that a reasonable adversary FTL (“The architecture of a storage system”section). will find no reason to keep holding him/her. Gracewipe 2) The PDE systems built on top of flash memory. This works as follows: During setup, the user selects three pass- type of PDE systems directly work on top of flash memory words, which can be used to derive the key for encrypting to provide deniability while handling the special nature of the corresponding volume: 1) hidden password that only flash. Table 1 Summary of PDE systems for desktop computers Scheme Method Adversary Deniable level Layer Scheme 1 (Anderson et al. 1998) Steganography Multi-snapshot Arbitrary File system Scheme 2 (Anderson et al. 1998) Steganography Single-snapshot Arbitrary File system StegFS (McDonald and Kuhn 2000) Steganography Single-snapshot Arbitrary File system StegFS (Pang et al. 2003; 2004) Steganography Multi-snapshot Arbitrary File system StegHide (Zhou et al. 2004) Steganography Multi-snapshot Arbitrary File system TrueCrypt (2012) Hidden volume Single-snapshot One Block device HIVE (Blass et al. 2014) ORAM Multi-snapshot Multiple Block device DataLair (Chakraborti et al. 2017) ORAM Multi-snapshot Multiple Block device DetWoORAM (Roche et al. 2017) ORAM Multi-snapshot Multiple Block device ECD (Zuck et al. 2017) Hidden volume Multi-snapshot One Block device Zhang et al. Cybersecurity (2018) 1:1 Page 9 of 20 PDE systems on top of block devices. Skillen et al. fast mode switching, MobiHydra introduces a special par- (Skillen and Mannan 2013, 2014; Skillen 2013)designed tition called shelter volume on the external storage, which Mobiflage, the first PDE scheme for mobile devices. They is used as a temporary storage partition which can tem- provided two versions of implementations: one in external porarily store the sensitive data being created in the pub- storage for FAT32 file system (Skillen and Mannan 2013), lic mode, without the need of entering the PDE mode and the other in internal storage for modified EXT4 file for storing hidden sensitive data. The data stored in the system (Skillen and Mannan 2014). The main contribu- shelter volume will be immediately synchronized to the tion of Mobiflage lies in its first incorporation of hidden hidden volume when the hidden mode is entered, and volumes technique to the Android devices. It works as are then eliminated from the shelter volume. To avoid follows: First, it fills the external storage with random deniability compromise, the sensitive data stored in the data. Second, it creates two volumes, a public volume for shelter volume will be encrypted by a random key which storing public non-sensitive data, applications and set- is encrypted by the public key and stored in the shelter tings, and a hidden volume for storing sensitive data. volume. In addition, some dummy files are maintained in Correspondingly, there are two operation modes, a pub- the shelter volume and updated periodically to obfuscate lic mode and a PDE mode. The public mode is used to the writes of hidden sensitive data to the shelter volume. manage the public volume for daily use, providing stor- MobiHydra, however, cannot eliminate the assumption of age encryption without deniability. In this mode, the data requiring a physical or an emulated FAT32 SD card. are encrypted with a decoy key, which is derived from a To eliminate the limitations of Mobiflage and MobiHy- decoy password. The user will be asked to provide the dra, Chang et al. designed MobiPluto (Chang et al. 2015), decoy password during the booting in order to enter the a file system friendly PDE design. The basic idea of Mobi- public mode. The PDE mode is used to manage the hid- Pluto is introducing an additional software layer between den volume which stores sensitive data whose existence the PDE and the file system. This software layer should needs to be denied when being coerced. In this mode, the satisfy three requirements: 1) Its existence should not be data are encrypted with a true key, which is derived from an indication of PDE; 2) It should provide virtual vol- atruepassword. Theusershould provide thetruepass- umes to file systems, and any block-based file systems word during system boot to activate the PDE mode. The can be deployed on a virtual volume; 3) It should convert exact location of the hidden volume is derived from the non-sequential allocation from a file system to sequential true password. allocation in the underlying PDE. To build such a layer, The FAT32 version of Mobiflage (Skillen and Mannan MobiPluto uses thin provisioning (Thornber), because: 2013) is specially designed for external storage formatted First, thin provisioning has been implemented by dm- using FAT32 file system, which requires the support of a thin-pool module, which has been a well-established tool physical or an emulated FAT32 SD card. This is because, in Linux kernel; Second, thin provisioning can allow to the hidden volume is part of the public volume, and is create thin volumes, each of which can be used to deploy placed at the end of the disk. In other words, if an EXT- any block-based file systems. Third, thin provisioning can like file system is deployed for the public volume, the data transform non-sequential allocation on the thin volume to stored in the hidden volume may be easily overwritten sequential allocation on the underlying storage. By com- by the public data due to the nature of EXT file sys- bining thin provisioning and hidden volumes, MobiPluto tem’s random allocation. To eliminate the aforementioned is able to achieve a “file system friendly” PDE design. assumption, the EXT4 version of Mobiflage (Skillen and Chang et al. further improved the usability of MobiPluto Mannan 2014) modifies the EXT4 driver such that it can in their extended work (Chang et al. 2018)byintroduc- support a sequential inode allocator and can be deployed ing a fast switching mechanism and using NFC cards to for the public volume to avoid overwriting the hidden vol- store strong PDE passwords. For PDE systems on mobile ume. However, the modification of the EXT4 driver itself devices, fast switching is a desired feature. When a device may be an indication of the existence of PDE. owner faces an emergency and wants to collect sensi- Yu et al. proposed MobiHydra (Yu et al. 2014)to tive information, he/she needs to instantly switch the improved Mobiflage in three aspects: 1) It can mitigate device to the hidden mode. However, it needs more than a novel booting-time attack being faced by Mobiflage; 1 min for the prior mobile PDE systems to switch modes, 2) It can support multiple levels of deniability. 3) It sup- because a full device rebooting is usually required. Fast ports mode switching without rebooting. To mitigate switching mechanism eliminates the need for rebooting the booting-time attack, MobiHydra obfuscates the time the device and the switching time is reduced to less than required for a wrong password (i.e., an arbitrary password 10 s (Chang et al. 2018). Their idea is to restart only except the true and the decoy password) during booting the Android framework rather than the entire device, such that the adversary is not able to identify the existence significantly reducing the switching time. For the hidden of PDE by simply entering a wrong password. To support volume, a strong password is required to protect sensitive Zhang et al. Cybersecurity (2018) 1:1 Page 10 of 20 data. However, users tend to choose weak passwords, thus properties in YAFFS2 to provide deniability. Therefore, it security may not be ensured. To address this issue, they is incompatible with the flash-based block devices using proposed to use NFC cards to store strong passwords for FTL, the most popular form of flash storage being used the user. Their observation is that most of the modern in mobile devices nowadays. Second, it suffers from deni- mobile devices are equipped with NFC features. ability compromises (Jia et al. 2017). This is because, to MobiCeal (Chang et al. 2018) designed the first block- prevent data at lower security levels from overwriting the based PDE system for mobile devices which can defend data at higher security levels, it disables garbage collection against a multi-snapshot adversary. The fundamental idea at the lower security levels. The adversary can easily iden- of MobiCeal is to use dummy writes to obfuscate writes tify this abnormal behavior and suspect existence of PDE performed in hidden mode. In this way, even though the (Jia et al. 2017). adversary can have access to multiple snapshots of the DEFTL (Jia et al. 2017) incorporated PDE into FTL. block image, it cannot tell whether a write is issued by DEFTL also has two modes, a public and a hidden mode. the hidden mode or not. In addition, MobiCeal can sup- The deniability of DEFTL is achieved by using the data port multiple deniability levels. Finally, it identifies various (and theirbehavior) in thepublicmodetodenythe side channel leakage present in prior PDE schemes for data (and their behavior) in the hidden mode. Most mobile device and eliminates them. importantly, to prevent the data written in the public mode from over-writing the data written in the hidden PDE systems on top of flash memory. Flash memory mode, DEFTL carefully modifies the block allocation and has significantly different nature compared to mechanical garbage collection strategies in the FTL such that the two disks, e.g., flash memory is update unfriendly and vulner- modes can be “stealthily” isolated without being known able to wear (“Flash memory” section). All the aforemen- by the adversary. Specifically, the public volume will allo- tioned block-based PDEs unfortunately may suffer from cate flash blocks from the head of the block pool and the deniability compromises in the underlying flash storage hidden volume will allocate flash blocks from the tail of due to the handling of the unique nature of flash memory. the pool. In addition, garbage collection in the two modes This is because: The unique characteristics of flash mem- will be modified as: In the public mode, garbage collec- ory require a special internal management, which creates tion will be performed actively to fill the head of the pool; a different view of data in flash memory, independent of in the hidden mode, garbage collection will be performed the view on the block layer. By having access to the raw actively to fill the tail of the pool. This can avoid that flash, the adversary can obtain this different view, which the public mode has used all the blocks in the head and starts to use the blocks in the tail, over-writing the hidden may allow it to observe those unexpected “traces” of sensi- sensitive data. DEFTL also provides a few attacks on the tive data, whose existence needs to be denied. To eliminate existing PDE systems for mobile devices. the aforementioned deniability compromise, a few PDE Table 2 summarizes the existing PDE systems for mobile systems directly incorporate PDE into flash memory. devices. They either use hidden volumes or steganogra- Peters et al. (2015)introducedDEFY, adeniable phy, and provide one or multiple deniability levels when encrypted file system based on flash file system YAFFS2 facing single-snapshot or multi-snapshot adversaries. (Robust Flash Storage: YAFFS 2002). In DEFY, operations Schemes based on hidden volumes (Skillen and Mannan on a higher security level are indistinguishable from the 2013, 2014;Yuetal. 2014;Chang et al. 2015, 2018) operations on a lower security level. In addition, DEFY usually have low performance overhead, but they can can mitigate over-writes of hidden data in the higher secu- only defend against single-snapshot adversaries. DEFY rity level by taking advantage of special properties offered by a log-structured file system. DEFY however, suffers (Peters et al. 2015) relies on steganography to defend from a few limitations. First, it strongly relies on system against multi-snapshot adversaries, but the performance Table 2 Comparison of PDE systems for mobile devices Scheme Method Adversary Deniability level Layer Mobiflage-FAT32 (Skillen and Mannan 2013) Hidden volume Single-snapshot One Block device Mobiflage-EXT4 (Skillen and Mannan 2014) Hidden volume Single-snapshot One Block device MobiHydra (Yu et al. 2014) Hidden volume Single-snapshot Multiple Block device MobiPluto Chang et al. (2015, 2018) Hidden volume Single-snapshot Multiple Block device MobiCeal (Chang et al. 2018) Steganography Multi-snapshot Multiple Block device DEFY (Peters et al. 2015) Steganography Multi-snapshot Arbitrary File system DEFTL (Jia et al. 2017) Hidden volume Single-snapshot One Controller (FTL) Zhang et al. Cybersecurity (2018) 1:1 Page 11 of 20 overhead is high. MobiCeal (Chang et al. 2018)usesa However, once sensitive data are discarded, the data lightweight “dummy write” mechanism to defend against owner may want to permanently remove them. Protec- multi-snapshot with acceptable performance overhead. tion of data confidentiality requires securely disposing those data to prevent the adversary from recovering all or Discussions portion of them. This is achieved by using secure deletion. About randomness being used in hidden volumes Hard disk drives (HDDs) and NAND flash mem- ory dominate the storage media of personal computing technique. The idea of hidden volumes technique is to hide the encrypted hidden volume among the random- devices. However, they are completely different in nature, ness being filled initially. To make sure that the ciphertexts and hence different secure deletion approaches should be being generated are indistinguishable from random bits, used to eliminate data from them. In the following, we the random bits can be drawn from the same distribution summarize secure deletion approaches for HDDs-based as the ciphertext space by using the encryption function and flash memory-based storage systems, respectively. itself as the PRNG (Skillen and Mannan 2014). This can We also summarize the recent works which complement ensure that the ciphertexts and the randomness are from conventional secure deletion approaches by taking care of the same source. In addition, the encryption being used the impacts created by past existence of the deleted data. in the hidden volumes technique is full disk encryption (FDE), which treats each disk sector as an autonomous Secure deletion for HDDs-based storage systems unit and assigns sector-specific IVs for chaining modes An HDD is a magnetic medium which supports in- such as CBC and XTS (Skillen and Mannan 2014). This place updates. Therefore, in HDDs-based storage sys- can help eliminate the correlations among ciphertexts due tems, when a file block is updated (or deleted), its old to similarities in files (e.g., file heads). version can be simply replaced by its new version (or random noise) in the storage medium to achieve secure Transferring data between desktop computers and deletion. mobile devices. SincePDE canbedeployedineither Physical medium layer. Due to lack of semantics of file desktop computers or mobile devices, there is a possibility system, single-file sanitization is not feasible at the phys- that the data owner would like to exchange data (e.g., files) ical medium layer. To delete the data, a naivest way is to between these two different platforms. The data exchange overwrite/destroy all the data on the physical medium. should be conducted in such a way that deniability should Tools such as degaussers can be used to sanitize data on not be compromised. When the file is non-sensitive, it HDDs (Kisseletal. 2006). can be simply transferred from a desktop computer to a Controller layer. At the controller layer, there are sev- eral standardized interfaces that permit reading/writing mobile device (or vice versa) without any deniability con- of fixed-sized blocks. Similarly, there are no semantics of cern. This can be done by simply using Internet or direct file system in this layer, thus the controller must sani- copying via cable, when both the source device and the tize every block to achieve secure deletion. To delete the destination device are working in the public mode. When the file is sensitive whose existence needs to be denied, to data, sanitize commands and overwrite techniques are transfer it from a desktop computer to a mobile device (or widely used in HDDs, e.g., secure erase commands offered vice versa), there are two known options: 1) If the file is by both SCSI (Incits: Scsi storage interfaces 2016)and transferred via Internet, to allow the data owner to deny ATA (Team work systems: Advanced technology attach- the transferring of this file (the attacker can collude with ment 2017). These sanitization commands work like a Internet ISPs and identify this event), covert communica- button that erases all data on the device by exhaustively tions approaches (Frèche et al. 2017;Huetal. 2017)can be overwriting every block with zeros or ones. used. 2) If the file is transferred via direct copying, the file Block device layer. Reardon et al. (2013)proposeda canbereadfromthe source device andwrittento thedes- secure deletion approach targeting persistent storage. tination device, when both devices are working the hidden Their approach relies on encryption and key wrapping. mode. Since the attacker cannot capture the device work- They useakey disclosure graphtomodel theadversarial ing in the hidden mode (“The assumptions required by knowledge about key generation and wrapping history. In PDE” section), the deniability will not be compromised. addition, a small securely-deleting key-value map is used Once the file is successfully transferred, it will be pro- to discard encryption key of the data, achieving secure tected by the PDE system in the destination device. deletion. File system layer. When deleting files in the EXT2, there Ensuring confidentiality of the deleted data via is a sensitive attribute for files and directories to indicate secure deletion that secure deletion should be used. (Bauer and Priyantha 2001) provided a patch that implements this attribute. By PDE is used to ensure confidentiality of sensitive data marking a block as free, the patch passes the free block which are preserved in the personal computing devices. Zhang et al. Cybersecurity (2018) 1:1 Page 12 of 20 to a kernel daemon, which maintains a list of blocks that deletion interface, which can be used to overwrite data must be sanitized. If the free block stores data from a with zeros in the underlying file system. For MySQL, sensitive file, instead of returning it to the file system (Stahlberg et al. 2007)proposedanapproach to delete as an empty block, this free block will be added to the entries by overwriting them with zeros, and the trans- work queue. When the system is idle, the sanitization dae- action log is encrypted and can be securely disposed by mon runs asynchronously to perform sanitization over the deleting the encryption key. For SQLite (SQLite: Pragma work queue, allowing the user to achieve immediate file statements 2017), there is a compile-time option to enable deletion. a secure deletion feature that overwrites deleted records (Joukov and Zadok 2005)proposedafile system with zeros. extension, purgefs, which uses block-based overwrit- Table 3 summarizes the existing secure deletion ing when blocks are returned to the file system’s free approaches for HDDs-based storage systems. Those blocks list. It supports overwriting file data and metadata approaches are incorporated into different layers of a stor- for all files or just files marked as sensitive. (Joukov age system and rely on either overwriting or encryption and Zadok 2006) also proposed three secure deletion for secure deletion. They may also have different dele- approaches for EXT3. The first approach is called EXT3 tion granularity. Generally, the encryption-based secure basic, which securely deletes data (but not metadata) deletion approaches for HDDs-based storage systems once by overwriting it. The second approach is called (Reardon et al. 2013;Petersonetal. 2005;Zhaoand EXT3 comprehensive, which overwrites file data and Mannan 2015) have relatively low performance overheads metadata by a configurable overwriting scheme. Both since only keys need to be deleted to achieve secure the aforementioned two overwriting-based approaches deletion. can securely delete all the data or just files whose extended attributes include a sensitive flag. The third Secure deletion for flash memory-based storage systems approach is based on intercepting files deletion events, The aforementioned secure deletion approaches for i.e., unlinking and truncating a file. The file to be deleted HDDs-based storage systems rely on properties of hard is moved into a special secure deletion directory, and disks: magnetism-based and supporting in-place updates. a background user-level tool shred (Plumb 2010)will However, flash memory does not possess these proper- delete the files in the secure deletion directory at regular ties. Wei et al. (2011) performed a series of experiments intervals. to show that the deletion techniques that work well for Peterson et al. (2005) optimized secure deletion for HDDs may not work properly for NAND flash. As NAND versioning file systems using an all-or-nothing transform flash is not magnetism-based, the degaussing method may (AONT). Using AONT, each data block is extended into damage NAND flash chips and render data unreadable, an encrypted data block along with a small stub. If any part but all the data may still remain intact (Wei et al. 2011). of the ciphertext is deleted, the entire message can not be Wei et al. (2011) created 1000 small files on an SSD, decrypted any more. In this way, a specific version of a file then dismantled the drive, and searched for the content can be quickly deleted by simply overwriting all the stubs. of those files. They found that the SSD contained up to In addition, to delete a large log file to which data have 16 stale copies of the tested files. This is because the been appended only, securely deleting all the blocks in its FTL creates redundant file copies during garbage collec- most recent version will achieve secure deletion on all its tion and out-of-place updates, which unfortunately will past versions. complicate the secure deletion design for flash memory. Application layer. The application layer can only inter- They also tested 13 single-file overwriting-based saniti- act with file system through a POSIX-compliant interface. zation tools (LSoft Technologies Inc: Active@ KillDisk A user-level application can securely erase all the data 2017; GEEP EDS LLC: Darik’s Boot and Nuke 2017), to on the storage medium by invoking a secure erase com- find out whether they work correctly for flash memory. mand (Incits: Scsi storage interfaces 2016;Teamwork Unfortunately, all those tools were not able to sanitize data systems: Advanced technology attachment 2017)inthe from flash memory: between 4 and 75% of the files’ con- hardware controller’ interface. A few files overwriting tent remained in the SATA SSDs, and between 0.57 and tools, e.g, srm (Jagdmann 2015)and wipe (Durak 2006), 84.9% of the data remained in USB drives. All the afore- can be used to securely remove files. Gracewipe (Zhao mentioned results indicate that securely deleting data and Mannan 2015), as has been discussed in “PDE for from NAND flash is much more challenging compared desktop computers” section, can achieve secure and veri- to HDDs, due to the special nature of NAND flash. In fiable deletion of encryption keys through a special dele- the following, we summarize the existing works aiming to tion password by taking advantage of TPM and Intel securely remove data from flash-based storage systems, TXT, thus making the encrypted data permanently inac- which are divided into two categories: overwriting-based cessible. In the case of database, there is also a secure and encryption-based. Zhang et al. Cybersecurity (2018) 1:1 Page 13 of 20 Table 3 Comparison of features offered by different secure deletion approaches for HDDs-based storage systems Scheme Method Layer Deletion granularity Degaussers (Kissel et al. 2006) Degaussing Physical medium The entire storage medium Secure erase commands (2016, 2017) Overwriting-based Controller The entire storage medium Reardon et al. (Reardon et al. 2013) Encryption-based Block device Single file Bauer et al. (Bauer and Priyantha 2001) Overwriting-based File system (EXT2) Single file Purgefs (Joukov and Zadok 2005) Overwriting-based File system (EXT2) Single file Joukov et al. (Joukov et al. 2006) Overwriting-based File system (EXT3) Single file Peterson et al. (Peterson et al. 2005) Encryption-based File system Single file File overwriting tools (Jagdmann 2015;Durak 2006) Overwriting-based Application Single file Gracewipe (Zhao and Mannan 2015) Encryption-based Application The entire storage medium Stahlberg et al. (Stahlberg et al. 2007) Based on overwriting Application (for database) Single entry/record and SQLite (SQLite: Pragma statements 2017) and encryption Overwriting-based secure deletion for NAND flash. some others, it does not cause any errors. They intro- A common idea for the overwriting-based secure deletion duced scrub budget, which refers to the number of times approaches is to replace the deleted data with meaningless that the NAND flash can allow to be scrubbed with- information, e.g., noisy random data. out exhibiting a significant risk of data errors. When the Physical medium layer. Similar to the physical medium scrub budget for a block is exceeded, secure deletion will layer in the HDDs-based storage systems (“Secure dele- be instead performed by other approaches (e.g., invok- tion for HDDs-based storage systems”section), data can ing garbage collection). More recently, Qin et al. (2013) notbedeleted in this layerdue to lack of semanticsof incorporated RAID-5 architecture to enhance the reliabil- the file system. To realize secure deletion in this layer, the ity and eliminate the negative effect of reprogram on flash entire flash chip should be destroyed. memory. Controller layer. A main type of flash controller is using File system layer. Sun et al. (2008) proposed a secure FTL to handle the special nature of NAND flash and to deletion method in YAFFS by investigating characteris- provide a block access interface to upper layers (“The tics of NAND flash memory. They proposed two secure architecture of a storage system” section). To securely deletion approaches, zero overwrite (similar to the scrub- delete data, the simplest way is to erase the correspond- bing (Wei et al. 2011)) and block erase. Especially, they ing flash blocks in the FTL (i.e., block erasure). However, define a costs-benefits model by comparing the overwrite erasures can only be performed in terms of flash blocks cost on the deleted pages and the erase cost on the block (“Flash memory” section). This will be overkill if only a that contains the deleted pages. Additionally, a new adap- portion of data being stored in a flash block needs to be tive hybrid scheme is applied to select the cheaper one deleted. Considering content of a file may be distributed between the two secure deletion approaches. Another in different pages of different flash blocks, sanitizing a file kernel-level zero-overwriting secure deletion approach using block erasure will be unavoidably expensive. was also proposed by Reardon et al (2012). Wei et al. proposed scrubbing (Wei et al. 2011)to By adding a new communication channel between the address the aforementioned issue. As programming indi- file system and the device driver, the file system can vidual pages is possible, the idea of scrubbing is to re- inform the device that particular blocks are no longer program the page, where the data should be securely valid, e.g., Trim (Intel Corporation: Intel Solid-State Drive deleted, to turn all its remaining ‘1’ bits into ‘0’. Note that Optimizer 2009) command and TrueErase (Diesburg et al. flash memory allows to individually program bit ‘1’ to ‘0’, 2012). With the information of invalid blocks, the device but the reverse operation is not feasible except perform- driver can implement its own efficient secure deletion ing a block erasure. A major concern of scrubbing is that without requiring data blocks to be explicitly overwrit- it may result in undefined behaviors due to possibility of ten by the file system. Especially, TrueErase is designed introducing read errors. To handle this concern, Wei et al. for the blocks belonging to files specifically marked as examined error rates for different types of flash memory sensitive. and showed that the error rates vary widely. For some Application layer. Since the application layer cannot flash devices, scrubbing causes frequent errors, while for directly touch the lower layers, secure deletion can only Zhang et al. Cybersecurity (2018) 1:1 Page 14 of 20 be achieved by filling the entire remaining free space. approach to perform standard data sanitization which can Reardon et al. (2012) proposed two user-level filling-based satisfy government agencies’ requirements (NSA/CSS and secure deletion approaches for YAFFS: purging and bal- DoD 5220.22-M) for the secure deletion. looning. Their basic idea is to fill the entire free space of Reardon et al. (2012)proposeddatanodeencrypted file the file system, such that all unused blocks of the physi- system (DNEFS), which enables secure data deletion for cal medium will no longer contain sensitive information. flash memory. They also incorporate DNEFS into flash By completely filling the file system’s empty space with file system UBIFS (Memory Technology Devices: UBIFS noise, all previously data deleted by the user are guaran- 2015). In DNEFS, they divide the entire flash memory into teed to have been erased. Compared to purging which two areas: a small key storage area and a large main data ensures rapid secure deletion of data from user-space, bal- storage area. They encrypt each data node (i.e., the unit of looning achieves a probabilistic continuous secure dele- I/O) with a unique key, and collocate the keys in the key tion guarantee by reducing the block reallocation period. area. Secure deletion is achieved by removing keys, which Braga et al. (2014) proposed two user-level approaches to can be performed efficiently, as keys are condensed in a securely delete files on Android smart phones. The one small area. that is designed to delete unencrypted files is also based DEFY (Peters et al. 2015) also provides secure dele- on filling. The cost of the filling-based secure deletion tion, complementary to its deniability. It leverages all- approaches is proportional to the size of free space avail- or-nothing transform (AONT), a cryptographic function able on the physical medium. A larger size of free space which can ensure that a missing portion of a message will will lead to a higher overhead in order to fill it (Reardon render the entire message irrecoverable. In this way, DEFY et al. 2013). The efficiency can be improved by perpetually can efficiently achieve secure deletion by only removing maintaining the free space of the physical medium within a small portion of the data being deleted. Braga et al. a limited range (Reardon et al. 2012). (Braga and Colito 2014) proposed two secure deletion approaches for Android phones, one of which is based on Encryption-based secure deletion for NAND flash. encryption. They modified the key management of EncFS Data can be rendered inaccessible by encrypting them (Wang et al. 2012), an encrypted file system, to ensure that and deleting the corresponding key. Boneh et al. pro- every file is encrypted with a unique key and a random posed the first encryption-based solution that securely IV. The removal of the unique key and IV makes the file deletes encrypted data stored on the tape by deleting the irrecoverable. cryptographic keys (Boneh and Lipton 1996). Table 4 summarizes the features offered by different secure deletion approaches for flash-based storage sys- Controller layer. Reliably destroying keys is challenging, tems, including overwriting-based and encryption-based as side-channel attacks based on semiconductor memory approaches. These approaches may also have differ- data remnants (Halderman et al. 2009) may allow an ent deletion granularity. Generally, the scrubbing-based attacker to recover the key or key-related information. secure deletion approaches (Wei et al. 2011;Qinet al. Swanson et al. (2010)proposedscrambleand finally erase (SAFE), which combines encryption and erasure tech- 2013;Sun et al. 2008; Reardon et al. 2012), which only niques to provide almost instant secure deletion with ver- targets at the invalid data, and the encryption-based ifiability. SAFE relies on the assumptions that data in the secure deletion approaches (Lee et al. 2008, 2010, 2011; SSDs are stored encrypted and the SSDs implement best Reardon et al. 2012), which only need to delete keys, practices of key management (e.g., the keys should never have relatively low performance overheads compared to leave the controller). It works as follows: Upon receiving the overwriting-based approaches. A major advantage of a sanitize command, it erases the controller’s key mem- encryption-based approaches is that, deleting a small key ory, such that the driver is not able to encrypt/decrypt the usually can be much more efficiently achieved compared data. It then erases every block on the device, and writes to deleting data which are large in size. all the pages with a known pattern, and erases them again. Finally, it reinitializes the device and performs a low-level A new direction for secure deletion format operation on the drive, and provides a new key to Secure deletion is used to securely dispose data once they the controller. become obsolete. This requires a security deletion guaran- File system layer. Lee et al. (2008, 2010)proposeda tee that the adversary should neither recover the deleted secure deletion approach for YAFFS, a log structured data, nor learn anything about them (“Secure deletion” NAND flash file system. By modifying YAFFS, they section). The question is, can we achieve the secure dele- encrypt files and force all keys of a specific file to be tion guarantee by simply deleting the data themselves? stored in the same block. Therefore, only one erase oper- The answer is unfortunately no. Conventional secure dele- tion approaches rely on either overwriting or encryption ation needs to be performed in order to securely delete a to make the deleted data inaccessible. However, the past file. Lee et al. (2011) then extended the aforementioned Zhang et al. Cybersecurity (2018) 1:1 Page 15 of 20 Table 4 Comparison of features offered by secure deletion approaches for flash-based storage systems Scheme Method Layer Deletion granularity Scrubbing (Wei et al. 2011) and SmSD (Qin et al. 2013) Overwriting-based Controller (FTL) A physical page Sun et al. (Sun et al. 2008) Overwriting-based File system (YAFFS) Single file Zero overwriting (Reardon et al. 2012) Overwriting-based File system (YAFFS) Single file Trim (Intel Corporation: Intel Solid-State Drive Optimizer 2009)and TrueErase (Diesburg et al. 2012) Overwriting-based File system Single file to device driver Purging and ballooning (Reardon et al. 2012) Overwriting-based Application All the invalid data Braga et al. (Braga and Colito 2014) for Overwriting-based Application All the invalid data unencrypted files SAFE (Swanson and Wei 2010) Based on overwriting Controller The entire and encryption storage medium Lee et al. (2008, 2010) Encryption-based File system (YAFFS) Single file Lee et al. (2011) Encryption-based File system (YAFFS) Single file DNEFS (Reardon et al. 2012) Encryption-based File system (UBIFS) Single file DEFY (Peters et al. 2015) Encryption-based File system (YAFFS) Single file Braga and Colito (2014) for Encryption-based File system (EncFS) Single file encrypted files existence of the deleted data may leave artifacts in the lay- structures. This also indicates that, although node 2 have out at all layers of a computing system (Bajaj and Sion been deleted, its structural artifacts remain in the data 2013b;Chenand Sion 2015) or create side effects on organization. Therefore by having access to tree T ,the the other data which have not been deleted (Bajaj and adversary may suspect the past existence of the sensitive Sion 2013a), and the adversary can potentially take advan- data which have been deleted, and tries to partially or fully tage of those structural artifacts or side effects to learn recover them. sensitive information about the deleted data (Bajaj and We also describe some concrete attack scenarios which Sion 2013a, b;Chenand Sion 2015). can take advantage of structural artifacts (Jia et al. 2016; To justify the impact of the past existence of the deleted Chen et al. 2016). Commodity NAND flash-based block data, we use a balanced binary search tree (BST) (Chen devices usually adopt a log-structured writing technique, et al. 2016). We first create a balanced BST by inserting in which flash blocks as well as pages within a block five nodes in the order of 2, 11, 13, 14, 1, and obtain tree are allocated sequentially (Min et al. 2012). As shown in T (Fig. 3a). We then delete node 2, obtaining tree T Fig. 4,A,B,C,and Dare thedatabeing writtentoNAND 1 2 (Fig. 3b). However, if we directly create the balanced BST flash, and each occupies a flash page. To securely delete by inserting nodes in the order of 11, 13, 14, 1, we will C, the user has two options: 1) We perform a scrubbing obtain tree T (Fig. 3c). This example indicates that, due (Wei et al. 2011) over the corresponding flash page. The to the past existence of node 2, T and T have different scrubbing technique however, will convert this page to a 2 3 Fig. 3 An example from (Chen et al. 2016) showing why structural artifacts matter. T , T ,and T are balanced BSTs. a Tree T , b Tree T ,and c Tree T 1 2 3 1 2 3 Zhang et al. Cybersecurity (2018) 1:1 Page 16 of 20 Fig. 4 Scrubbing-based secure deletion page with all “0” bits (i.e., a zero page). By having access flash with a full scrubbing (Wei et al. 2011)orablockera- to the storage state after deletion of C, the adversary will sure may provide the adversary a clue that there was a notice the zero page and suspect that a past deletion has deletion in the past, Jia et al. proposed to perform a par- been performed on it. 2) We encrypt A, B, C, and D tial scrubbing, i.e., partial page reprogramming and partial with different keys, and delete key for C (Reardon et al. block erasure, to only modify a portion of the bits in the 2012). However, the adversary can also find out that C has page/block storing the deleted data, avoiding producing a been deleted since it cannot be successfully decrypted to zero page or an all-“1” block. plaintexts which are semantically meaningful. A further Truly secure deletion. NFPS (Jia et al. 2016)aimed to attack will be performed to fully or partially recover C conceal the past existence of the deleted data in the by taking advantage of the correlation between C and its NAND flash memory. However, it still cannot sanitize the neighboring data (Chen et al. (2016) provides a concrete structural artifacts introduced by the deleted data. Chen attack scenario which can take advantage of the structural et al. (2016) investigated another novel security notion, artifacts to recover bitcoin transactions). namely, truly secure deletion, which can ensure the saniti- To remove impacts of the past existence of the deleted data (including both structural artifacts and side effects), zation of both the data and the structural artifacts. a few recent works investigate undetected secure dele- To achieve truly secure deletion, Chen et al. (2016)pro- tion, truly secure deletion, history independence, as well posed TedFlash, a truly secure deletion scheme for NAND as untraceable deletion. flash-based block devices. In TedFlash, the data of every write to NAND flash will be placed to an empty loca- Undetectable secure deletion. Jia et al. (2016)tried to tion which is randomly selected. Note that as the random achieve the secure deletion guarantee by hiding the dele- placement of data is independent and does not affect tion history. Intuitively, if the deletion history is concealed theplacementsofany otherdata, TedFlash canelimi- from the adversary, he/she should not be able to find nate the structural artifacts brought by each write. Most out whether there was a deletion in the past, eliminating importantly, the random placement technique is exclu- his/her possibility in recovering the deleted data or learn- sively feasible for NAND flash, because: 1) Random seeks ing anything about the deleted data. Based on this key on flash memory are as efficient as sequential seeks. 2) observation, Jia et al. investigated a novel security notion The random placements can distribute data evenly among for NAND flash-based block devices, i.e., undetectable flash, naturally achieving a good wear leveling. secure deletion, to achieve two security properties: 1) Data to be deleted are completely removed from NAND flash History independence. History independence is pro- memory, which ensures that the adversary cannot have posed to prevent historic information about the pattern of access to thedataoncetheyhavebeendeleted;2)The access to a data structure from being leaked through its deletion history is concealed from the adversary, which representation when observed by an adversary (Chen and ensures that the adversary cannot gain any knowledge Sion 2015). History independence ensures that by hav- about whether there was a deletion in the past. ing access to a storage state, the adversary is not able to To realizeundetectablesecuredeletion, Jiaetal. pro- identify the operation sequence which leads to this state. posed NAND Flash Partial Scrubbing (NFPS), an unde- In other words, given two operation sequences leading to tectable secure deletion scheme for NAND flash-based thesamestorage state: onesequencehas adeleteopera- block devices. Having observed that deleting data from tion (e.g., delete D) and its corresponding insert operation Zhang et al. Cybersecurity (2018) 1:1 Page 17 of 20 (e.g., insert D), and the other sequence does not have the Untraceable deletion. Bajaj and Sion (2013a)introduced aforementioned delete and insert operation, the adversary untraceable secure deletion, aiming to remove side effects will not be able to differentiate which operation sequence of the data being deleted (i.e., the impacts of the past led to this storage state. Therefore, history independence existence of the deleted data on the data which are still ensures that after having removed a data record, the stor- preserved in the computing devices). The corresponding age state is somehow equivalent to a state that the deleted design, Ficklebase, is to achieve untraceable secure dele- record never exists, naturally achieving the secure dele- tion in the context of relational databases. In Ficklebase, tion guarantee. Implicitly, history independence ensures once a tuple is expired, all its side-effects will be removed that no structural artifacts will be introduced. Other- via versioning and query rewriting. wise, the memory representation of each storage state Table 5 summarizes the secure deletion approaches cannot be “independent” of the operation sequence which handle impacts of the past existence of deleted leading to it. data, including both structural artifacts and side effects. Bajaj and Sion (2016; Bajaj and Sion 2013b)designed The approaches can also securely remove data themselves, history independent file system (HIFS), the first approach and may be incorporated into different layers of a storage which can provide history independence for file storage system. over mechanical hard disks. HIFS uses a history indepen- dent hash table to allocate file data to the underlying block Future directions device in a history independent manner. Most signifi- This section provides an overview of the promising cantly, they can simultaneously achieve history indepen- research directions for both PDE and secure deletion. dence and preserve locality to improve I/O performance. However, HIFS does not work for flash-based storage Achieving plausible deniability and secure deletion systems (“The architecture of a storage system”section), in a single system. Most of the existing systems either since flash memory usually has its own internal software provide deniability or achieve secure deletion. However, layer (e.g., FTL) which is used to transparently handle its data confidentiality should be simultaneously ensured special characteristics by using a special history depen- during and after the lifetime of the data, because: First, dent block placement technique. by recovering the data being deleted, the adversary can Chen et al. thus proposed HiFlash (Chen and Sion 2015), achieve a similar gain comparable to compromising con- aiming to achieve history independence in flash-based fidentiality of the data being stored; Second, if the con- block devices. To achieve history independence, HiFlash fidentiality of the data cannot be ensured during their enforces a bijection between block device and flash mem- lifetime, secure deletion (i.e., ensuring confidentiality of ory. Specifically, HiFlash always places data records, which the data after their lifetime) turns meaningless since are written to the same block-device locations, to the the adversary has already obtained the data before they same flash locations, regardless of their write patterns. are “securely” removed. Therefore, we expect a system However, by introducing a bijective mapping between which can achieve both plausible deniability and secure block device and flash memory, HiFlash can only remove deletion. Simply combining the existing PDE and secure structural artifacts introduced by the software component deletion may be problematic, since secure deletion may staying between the block device and the flash mem- require a fine-grained encryption mechanism, and plau- ory (e.g., FTL). It is unfortunately not able to remove sibly deniable encryption is not necessarily designed as the structural artifacts introduced from the upper layers. fine-grained. In addition, when pre-processing data for Therefore, it strongly relies on the assumption that the secure deletion purpose, careful consideration may be upper layer has eliminated the structural artifacts, which needed to avoid bringing in any deniability compromises. is not necessarily true. Theonlyattempt forthistypeofsystemisDEFY(Peters Table 5 Comparison of features offered by secure deletion approaches designed to handle impacts created by past existence of the deleted data NFPS (Jia et al. 2016) TedFlash (Chen HIFS (Bajaj and HiFlash (Chen and Sion 2015) Ficklebase (Bajaj and Sion 2013a) et al. 2016) Sion 2013b; Bajaj et al. 2016) Method to Overwriting Encryption Overwriting Overwriting Encryption sanitize data Method to Partial scrubbing Random History indepen- One-one mapping Versioning, query rewriting sanitize past placement dent hash table impacts technique Layer Controller (FTL) Controller (FTL) File system Controller (FTL) Application Zhang et al. Cybersecurity (2018) 1:1 Page 18 of 20 et al. 2015), but DEFY itself may suffer from deniability Authors’ contributions All authors read and approved the final manuscript. compromise (Jia et al. 2017). Competing interests Providing confidentiality guarantee for light-weight The authors declare that they have no competing interests. computing devices. Computing devices nowadays are turning more and more light-weight. Wearable devices Publisher’s Note like smart watches and smart glasses, IoT devices like Springer Nature remains neutral with regard to jurisdictional claims in smart home hubs or smart plugs, are increasingly popular published maps and institutional affiliations. today. Those light-weight computing devices are usually equipped with limited computational power. However, Author details both the PDE and the secure deletion usually require Data Assurance and Communication Security Research Center, Chinese Academy of Sciences, Beijing, China. State Key Laboratory of Information expensive encryption operations, and thus cannot directly Security, Institute of Information Engineering, Chinese Academy of Sciences, fit the use of light-weight devices. This can be mitigated 3 Beijing, China. School of Cyber Security, University of Chinese Academy of by either outsourcing part of the expensive computation Sciences, Beijing, China. School of Information Systems, Singapore Management University, Singapore, Singapore. Department of Computer to the third-party cloud providers (without confidentiality Science, Michigan Technological University, Houghton, USA. compromise) or reducing the level of security to improve performance. Received: 19 January 2018 Accepted: 17 April 2018 Eliminating deniability compromise and data leakage. The existing PDE/secure deletion systems mainly focus References on external storage, and may neglect the security leakage Amazon: New SSD-Backed Elastic Block Storage (2016). Retrieved December in other sources like memory. By having obtained snap- 17, 2017, from https://aws.amazon.com/blogs/aws/new-ssd-backed- elastic-block-storage/ shots of the victim computing device, the adversary may Anderson R, Needham R, Shamir A (1998) The steganographic file system. capture the memory state, and may compromise deni- In: International Workshop on Information Hiding. Springer. ability or recover the data being deleted by performing pp 73–82 forensic analysis. Therefore, processing of the sensitive Apple: Use filevault to encrypt the startup disk on your mac (2017). Retrieved December 17, 2017, from https://support.apple.com/en-us/HT204837 data should be conducted in an isolated memory region, Assange J, Weinmann RP, Dreyfus S (2012) Rubberhose: Cryptographically which cannot be learned by the adversary. This isolated deniable transparent disk encryption system. Retrieved December 17, memory region can be created by using trusted execution 2017, from https://web.archive.org/web/20120716034441/http:// marutukku.org/ environment (e.g., Intel SGX (Intel: Intel software guard Bajaj S, Chakraborti A, Sion R (2016) Practical foundations of history extensions 2017)). independence. Information Forensics and Security. IEEE Transactions on 11(2):303–312 Secure against quantum computing. Both PDE and Bajaj S, Sion R (2013) Ficklebase: Looking into the future to erase the past. In: Data Engineering (ICDE) 2013 IEEE 29th International Conference on. secure deletion may rely on encryption (e.g., AES-128, IEEE. pp 86–97 XTS-AES), which is not necessarily quantum resistant. Bajaj S, Sion R (2013) Hifs: History independence for file systems. The development of quantum computers seems to be In: Proceedings of the 2013 ACM SIGSAC conference on Computer & communications security. ACM. pp 1285–1296 expedited recent years (Wikipedia: Timeline of quan- Ballard Spahr LLP: State and local governments move swiftly to sue equifax tum computing 2018). Therefore, there will be a need (2017). Retrieved December 17, 2017, from https://www.natlawreview. to ensure that the PDE and the secure deletion are com/article/state-and-local-governments-move-swiftly-to-sue-equifax Bauer S, Priyantha NB (2001) Secure data deletion for linux file systems. In: secure against quantum computing, which requires care- Usenix Security Symposium. Volume 174 fully checking the existing cryptography primitives being Blass EO, Mayberry T, Noubir G, Onarlioglu K (2014) Toward robust hidden used and replacing those which are vulnerable to quantum volumes using write-only oblivious ram. In: Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security. ACM. computing with the ones that are quantum resistant. pp 203–214 Boneh D, Lipton RJ (1996) A revocable backup system. In: USENIX Security Conclusion Symposium. pp 91–96 In this survey, we summarize the existing approaches Braga A, Colito A (2014) Adding Secure Deletion to an Encrypted File System on Android Smartphones. In: SECURWARE 2014, The Eighth International for protecting data confidentiality against snapshot adver- Conference on Emerging Security Information, Systems and Technologies. saries. Our survey covers techniques for both PDE and pp 106–110 secure deletion, which can ensure data confidentiality Breeuwsma M, De Jongh M, Klaver C, Van Der Knijff R, Roeloffs M (2007) Forensic data recovery from flash memory. Small Scale Digit Device during and after their lifetime, respectively. We also dis- Forensic J 1(1):1–17 cuss a few promising future directions. Cai Y, Luo Y, Ghose S, Mutlu O (2015) Read disturb errors in mlc nand flash memory: Characterization, mitigation, and recovery. In: Dependable Systems and Networks (DSN), 2015 45th Annua IEEE/IFIP International Acknowledgment Conference on. IEEE. pp 438–449 This work was partially supported by the National Key Research & Development Canetti R, Dwork C, Naor M, Ostrovsky R (1997) Deniable encryption. In: Annual Program of China (Grant No. 2017YFC0822704) and National Natural Science International, Cryptology Conference. Springer. pp 90–104 Foundation of China (No. 61602476, No. 61772518 and No. 61602475). Zhang et al. Cybersecurity (2018) 1:1 Page 19 of 20 Cbsnews: Apple’s celebrity icloud leak probably has mundane causes (2014). Intel Corporation: Intel Solid-State Drive Optimizer (2009). Retrieved December Retrieved December 17, 2017, from https://www.cbsnews.com/news/ 17, 2017, from http://download.intel.com/design/flash/nand/mainstream/ apples-celebrity-icloud-leak-probably-has-mundane-causes/ IntelSSDOptimizerWhitePaper.pdf Chakraborti A, Chen C, Sion R (2016) Poster: Datalair: A storage block device Jagdmann D (2015) srm - Linux man page. Retrieved December 17, 2017, from with plausible deniability. ACM https://www.mankier.com/1/srm Chakraborti A, Chen C, Sion R (2017) Datalair: Efficient block storage with Jia S, Xia L, Chen B, Liu P (2016) Nfps: Adding undetectable secure deletion to plausible deniability against multi-snapshot adversaries. Proc Priv Enhanc flash translation layer. In: Proceedings of the 11th ACM on Asia Conference Technol 3:175–193 on Computer and Communications Security. ACM. pp 305–315 Chang B, Cheng Y, Chen B, Zhang F, Zhu WT, Li Y, Wang Z (2018) User-friendly Jia S, Xia L, Chen B, Liu P (2017) Deftl: Implementing plausibly deniable deniable storage for mobile devices. Comput Secur 72:163–174 encryption in flash translation layer. In: Proceedings of the 24th ACM Chang B, Wang Z, Chen B, Zhang F (2015) Mobipluto: File system friendly conference on Computer and communications security. ACM deniable storage for mobile devices. In: Proceedings of the 31st Annual Joukov N, Papaxenopoulos H, Zadok E (2006) Secure deletion myths, issues, Computer Security Applications Conference. ACM. pp 381–390 and solutions. ACM Work Storage Secur Survivability:61–66 Chang B, Zhang F, Chen B, Li Y, Zhu WT, Tian Y, Wang Z, Ching A (2018) Joukov N, Zadok E (2005) Adding secure deletion to your favorite file system. Mobiceal: Towards secure and practical plausibly deniable encryption on In: Third IEEE International Security in Storage Workshop (SISW’05). mobile devices. In: Proceedings of the 48th IEEE/IFIP International IEEE. p 8 Conference on Dependable Systems and Networks Klonowski M, Kubiak P, Kutyłowski M (2008) Practical deniable encryption. Chen B, Jia S, Xia L, Liu P (2016) Sanitizing data is not enough!: towards SOFSEM 2008: Theory Pract Comput Sci:599–609 sanitizing structural artifacts in flash media. In: Proceedings of the 32nd Lee B, Son K, Won D, Kim S (2011) Secure data deletion for usb flash memory. Annual Conference on Computer Security Applications. ACM. pp 496–507 J Inf Sci Eng 27(3):933–952 Chen B, Sion R (2015) Hiflash: A history independent flash device. arXiv Lee C, Sim D, Hwang JY, Cho S (2015) F2fs: a new file system for flash storage. preprint arXiv:1511.05180 Usenix Conf File Storage Technol:273–286 Congress US (1996) HealthInsurancePortability and Accountability Act. Retrieved Lee J, Heo J, Cho Y, Hong J, Shin SY (2008) Secure deletion for nand flash file November 9, 2017, from http://www.hhs.gov/ocr/privacy/index.html system. In: Proceedings of the 2008 ACM symposium on Applied Congress US (1999) Gramm-Leach-Bailey Act. Retrieved November 9, 2017, computing. ACM. pp 1710–1714 from http://www.gpo.gov/fdsys/pkg/PLAW-106publ102/pdf/PLAW- Lee J, Yi S, Heo J, Park H, Shin SY, Cho Y (2010) An efficient secure deletion 106publ102.pdf scheme for flash file systems. J Inf Sci Eng 26(1):27–38 Czeskis A, Hilaire DJS, Koscher K, Gribble SD, Kohno T, Schneier B (2008) LSoft Technologies Inc: Active@ KillDisk (2017). Retrieved December 17, 2017, Defeating encrypted and deniable file systems: Truecrypt v5. 1a and the from http://www.killdisk.com/ case of the tattling os and applications. In: HotSec McDonald AD, Kuhn MG (2000) Stegfs: A steganographic file system for linux. Durak B (2006) wipe - Linux man page. Retrieved December 17, 2017, from In: Information Hiding. Springer. pp 463–477 https://linux.die.net/man/1/wipe Memory Technology Devices: UBIFS (2015). Retrieved December 17, 2017, Diesburg S, Meyers C, Stanovich M, Mitchell M, Marshall J, Gould J, Wang AIA, from http://www.linux-mtd.infradead.org/doc/ubifs.html Kuenning G (2012) Trueerase: Per-file secure deletion for the storage data Meng B, Wang Jq (2009) A receiver deniable encryption scheme. In: path. In: Proceedings of the 28th, Annual Computer Security Applications International Symposium on Information Processing (ISIP ’09) Conference. ACM. pp 439–448 Meng B, Wang J (2010) An efficient receiver deniable encryption scheme and Dürmuth M, Freeman DM (2011) Deniable encryption with negligible its applications. JNW 5(6):683–690 detection probability: An interactive construction. In: Annual International Microsoft: Bitlocker (2017). Retrieved December 17, 2017, from https://docs. Conference on the Theory and Applications of Cryptographic Techniques. microsoft.com/en-us/windows/device-security/bitlocker/bitlocker- Springer. pp 610–626 overview Frèche G, Bloch MR, Barret M (2017) Polar codes for covert communications Min C, Kim K, Cho H, Lee SW, Eom YI (2012) Sfs: random write considered over asynchronous discrete memoryless channels. Entropy 20(1):3 harmful in solid state drives. In: FAST. p 12 Garfinkel SL, Shelat A (2003) Remembrance of data passed: A study of disk O’Neill A, Peikert C, Waters B (2011) Bi-deniable public-key encryption. Adv sanitization practices. IEEE Secur Priv 99(1):17–27 Cryptol–CRYPTO 2011:525–542 Geambasu R, Kohno T, Levy AA, Levy HM (2009) Vanish: Increasing data privacy Pang H, Tan KL, Zhou X (2003) Stegfs: A steganographic file system. In: Data with self-destructing data. In: USENIX Security Symposium. pp 299–316 Engineering, 2003. Proceedings. 19th International Conference on. IEEE. GEEP EDS LLC: Darik’s Boot and Nuke (2017). Retrieved December 17, 2017, pp 657–667 from http://www.dban.org/ Pang H, Tan KL, Zhou X (2004) Steganographic schemes for file system and b-tree. IEEE Trans Knowl Data Eng 16(6):701–713 Goldreich O, Ostrovsky R (1996) Software protection and simulation on oblivious rams. J ACM (JACM) 43(3):431–473 Perlman R (2005) The Ephemerizer: Making Data Disappear. In: Journal of Gutmann P (1996) Secure deletion of data from magnetic and solid-state Information System Security. Citeseer memory. In: Proceedings of the Sixth USENIX Security Symposium, San Perlman R (2005) File system design with assured delete. In: Third IEEE Jose, CA. Volume 14 International Security in Storage Workshop (SISW’05). IEEE. p 6 Halderman JA, Schoen SD, Heninger N, Clarkson W, Paul W, Calandrino JA, Peters TM, Gondree MA, Peterson ZN (2015) DEFY: A deniable, encrypted file Feldman AJ, Appelbaum J, Felten EW (2009) Lest we remember: cold-boot system for log-structured storage. In: 22th Annual Network and Distributed attacks on encryption keys. Commun ACM 52(5):91–98 System Security Symposium. NDSS Han J, Pan M, Gao D, Pang H (2010) A multi-user steganographic file system on Peterson ZN, Burns RC, Herring J, Stubblefield A, Rubin AD (2005) untrusted shared storage. In: Proceedings of the 26th Annual Computer Secure deletion for a versioning file system. In: FAST. Volume 5. Security Applications Conference. ACM. pp 317–326 pp 143–154 Howlader J, Basu S (2009) Sender-side public key deniable encryption scheme. Plumb C (2010) shred 1 - Linux man page. Retrieved December 17, 2017, from In: Advances in Recent Technologies in Communication and Computing, https://linux.die.net/man/1/shred 2009. ARTCom’09. International Conference on. IEEE. pp 9–13 Qin Y, Tong W, Liu J, Zhu Z (2013) Smsd: A smart secure deletion scheme for Hu J, Yan S, Zhou X, Shu F, Wang J (2017) Covert Communication in Wireless ssds. J Converg 4:30–35 Relay Networks. In: GLOBECOM 2017 - 2017 IEEE Global Communications Kissel R, Scholl M, S S, Li X (2006) Guidelines for Media Sanitization. Retrieved Conference. pp 1–6 December 17, 2017, from http://ws680.nist.gov/publication/get_pdf.cfm? Ibrahim MH (2009) Receiver-deniable public-key encryption. IJ Netw Secur pub_id=50819 8(2):159–165 Reardon J, Basin D, Capkun S (2013) Sok: Secure data deletion. In: Security and Incits: Scsi storage interfaces (2016). Retrieved December 17, 2017, from http:// Privacy (SP), 2013 IEEE Symposium on. IEEE. pp 301–315 www.t10.org/index.html Reardon J, Capkun S, Basin D (2012) Data node encrypted file system: Efficient Intel: Intel software guard extensions (2017). Retrieved December 19, 2017, secure deletion for flash memory. In: Proceedings of the 21st USENIX from https://software.intel.com/en-us/sgx conference on Security symposium, USENIX Association. pp 17–17 Zhang et al. Cybersecurity (2018) 1:1 Page 20 of 20 Reardon J, Marforio C, Capkun S, Basin D (2012) Secure deletion on Zhou X, Pang H, Tan KL (2004) Hiding data accesses in steganographic file log-structured file systems. ASIACCS 2012 system. In: Data Engineering, 2004. Proceedings. 20th International Reardon J, Ritzdorf H, Basin D, Capkun S (2013) Secure data deletion from Conference on. IEEE. pp 572–583 persistent media. In: ACM Sigsac Conference on, Computer & Zuck A, Shriki U, Porter DE, Tsafrir D (2017) Preserving hidden data with an Communications Security. pp 271–284 ever-changing disk. Work Hot Top Oper Syst Robust Flash Storage: YAFFS (2002). Retrieved December 17, 2017, from http:// www.yaffs.net/yaffs-overview Roche DS, Aviv A, Choi SG, Mayberry T (2017) Deterministic, stash-free write-only oram. In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security. ACM. pp 507–521 Sarbanes USP, Oxley URMG (2002) Sarbanes-Oxley Act. Retrieved November 9, 2017, from http://www.sec.gov/about/laws.shtml#sox2002 Skillen A. (2013) Deniable storage encryption for mobile devices. PhD thesis. Concordia University Skillen, A, Mannan M (2013) On implementing deniable storage encryption for mobile devices Skillen A, Mannan M (2014) Mobiflage: Deniable storage encryptionfor mobile devices. IEEE Trans Dependable Secure Comput 11(3):224–237 Sourceforge: FreeOTFE A free “on-the-fly” transparent disk encryption program for PC and PDAs (2017). Retrieved December 19, 2017, Project website: https://sourceforge.net/projects/freeotfe.mirror/ Sourceware: Jffs2 (2003). Retrieved November 9, 2017, from https://www. sourceware.org/jffs2/ SQLite: Pragma statements (2017). Retrieved December 17, 2017, from http:// www.sqlite.org/pragma.html#pragma_secure_delete Stahlberg P, Miklau G, Levine BN (2007) Threats to privacy in the forensic analysis of database systems. ACM SIGMOD conf Manag data:91–102 Stefanov E, Van Dijk M, Shi E, Fletcher C, Ren L, Yu X, Devadas S (2013) Path oram: an extremely simple oblivious ram protocol. In: Proceedings of the 2013 ACM SIGSAC conference on Computer & communications security. ACM. pp 299–310 Sun K, Choi J, Lee D, Noh SH (2008) Models and design of an adaptive hybrid scheme for secure deletion of data in consumer electronics. Consumer Electronics. IEEE Trans 54(1):100–104 Swanson S, Wei M (2010) Safe: Fast, verifiable sanitization for ssds. University of California-San Diego, San Diego, CA Tang Y, Lee PP, Lui J, Perlman R (2012) Secure overlay cloud storage with access control and assured deletion. Dependable and Secure Computing. IEEE Trans 9(6):903–916 Team work systems: Advanced technology attachment (2017). Retrieved December 17, 2017, from https://teamws.com/glossary/ata-advanced- technology-attachment/ Thornber J Thin provisioning tools. https://github.com/jthornber/thin- provisioning-tools Troncoso C, Diaz C, Dunkelman O, Preneel B (2007) Traffic analysis attacks on a continuously-observable steganographic file system. In: International Workshop on Information Hiding. Springer. pp 220–236 TrueCrypt: Free open source on-the-fly disk encryption software.version 7.1a (2012). Retrieved December 19, 2017, Project website: http://www. truecrypt.org/ Varun S, Shibin K, Anoop S, Vivek K (2017) Magikfs - The Steganographic Filesystem On Linux. Retrieved December 19, 2017, Project website: http:// magikfs.sourceforge.net/ Wang Z, Murmuria R, Stavrou A (2012) Implementing and Optimizing an Encryption Filesystem on Android. In: IEEE International Conference on Mobile Data Management. pp 52–62 Wei MYC, Grupp LM, Spada FE, Swanson S (2011) Reliably erasing data from flash-based solid state drives. In: FAST. Volume 11. pp 8–8 Wikipedia: Timeline of quantum computing (2018). Retrieved December 19, 2017, from https://en.wikipedia.org/wiki/Timeline_of_quantum_ computing Yu X, Chen B, Wang Z, Chang B, Zhu WT, Jing J (2014) Mobihydra: Pragmatic and multi-level plausibly deniable encryption storage for mobile devices. In: Information Security - 17th International Conference, ISC 2014, Hong Kong, China, October 12-14 2014. Proceedings. pp 555–567 Zarras A, Kohls K, Dürmuth M, Pöpper C (2016) Neuralyzer: Flexible expiration times for the revocation of online data. In: Proceedings of the Sixth ACM on Conference on Data and Application Security and Privacy. ACM. pp 14–25 Zhao L, Mannan M (2015) Gracewipe: Secure and verifiable deletion under coercion. In: Network and Distributed System Security Symposium

Journal

CybersecuritySpringer Journals

Published: Jun 5, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off