Plausible deniability (PD) allows at-risk users to deny the existence of their sensitive data stored on storage devices. This is critical to protect the privacy and the personal safety of users, as adversaries might force users to decrypt their devices, risking the disclosure of sensitive data that could endanger their lives and liberty. In this work, we show how current PD systems built on flash memory fail to obscure distinguishable data layouts created when hidden data is written. This deficiency makes them vulnerable to coercive adversaries who can capture single or multiple data snapshots of storage devices for scrutiny. To defend against this threat, we propose MUTE, a perMUTation-based PD systEm designed for flash memory. Building upon widely-adopted full disk encryption (FDE) mechanisms that provide device-level data encryption, MUTE modifies the distribution of initialization vectors (IV s) for encryption blocks within FDE, translating the hidden data into a permutation derived from the IV. Unlike other PD solutions, MUTE allows for storing hidden data without requiring the reduction of storage capacity. Moreover, it preserves the plausible deni-ability of the hidden data in a provably secure manner by maintaining the original logic of data operations on the flash memory without changing the data layout. We implement MUTE in the flash translation layer (FTL) of flash-based SSDs using FEMU, a widely-used emulator supporting flash mem-ory research. Our evaluation with various micro-benchmarks and real-world workloads demonstrates that MUTE provides practical write and read throughputs of 23.4 MB/s and 15.7 MB/s and a capacity of 25.3 GB for hidden data in a 512 GB SSD, comparable with existing PD systems. MUTE achieves strong PD guarantees for flash-based devices against coercive adversaries, outperforming current PD systems. Index Terms-Plausible Deniability, Data Layout, Flash-based SSDs

Deep learning models are vulnerable to privacy attacks due to their tendency to memorize individual training examples. Theoretically-sound defenses such as differential privacy can defend against this threat, but model performance often suffers. Empirical defenses may thwart existing attacks while maintaining model performance but do not offer any robust theoretical guarantees. In this paper, we explore a new strategy based on the concept of plausible deniability. We introduce a training algorithm called **P**lausibly **D**eniable **S**tochastic **G**radient **D**escent (PD-SGD). The core of this approach is a rejection sampling technique, which probabilistically prevents updating model parameters whenever a mini-batch cannot be plausibly denied. We provide theoretical results showing that PD-SGD effectively mitigates privacy leakage from individual data points. Experiments demonstrate the scalability of PD-SGD and the favorable privacy-utility trade-off it offers compared to existing defense methods.

There is a flurry of recent research papers proposing novel differentially private machine learning (DPML) techniques. These papers claim to achieve new state-of-the-art (SoTA) results and offer empirical results as validation. However, there is no consensus on which techniques are most effective or if they genuinely meet their stated claims. Complicating matters, heterogeneity in codebases, datasets, methodologies, and model architectures make direct comparisons of different approaches challengingIn this paper, we conduct a reproducibility and replicability (R+R) experiment on 11 different SoTA DPML techniques from the recent research literature. Results of our investigation are varied: while some methods stand up to scrutiny, others falter when tested outside their initial experimental conditions. We also discuss challenges unique to the reproducibility of DPML, including additional randomness due to DP noise, and how to address them. Finally, we derive insights and best practices to obtain scientifically valid and reliable results.

Automatic Speech Recognition (ASR) systems take audio signals as inputs and output the corresponding text transcriptions. The text is then used to execute commands and perform searches in several application domains, including security-critical applications such as smartphone assistants, smart home assistants, and self-driving car assistants. Signal processing attacks are one of the most recent types of attacks designed to fool ASR models. Signal processing attacks exploit the feature extraction stage of the ASR pipeline and add perturbations to the audio. These attacks are capable of generating wrong transcriptions of the audio signals even though the attacked audio sounds similar to the original audio. Existing defences for adversarial attacks are neural networks that act as a filter to remove attacks from audio waveforms. The heuristic-based training objective function used in training these filter networks has a negative impact on the performance. Also, there is a disconnect between the training objective function and the application objective function. We address these problems and propose a novel self-supervised fine-tuning algorithm to make existing ASR models robust to adversarial attacks. We do extensive experimentation on our method against signal processing attacks across four different scenarios, and in three out of four scenarios, our method exhibits the best results.

State of the art generative models of human-produced content are the focus of many recent papers that explore their use for steganographic communication. In particular, generative models of natural language text. Loosely, these works (invertibly) encode message-carrying bits into a sequence of samples from the model, ultimately yielding a plausible natural language covertext. By focusing on this narrow steganographic piece, prior work has largely ignored the significant algorithmic challenges, and performance-security tradeoffs, that arise when one actually tries to build a messaging pipeline around it. We make these challenges concrete, by considering the natural application of such a pipeline: namely, "dead-drop" covert messaging over large, public internet platforms (e.g. social media sites). We explicate the challenges and describe approaches to overcome them, surfacing in the process important performance and security tradeoffs that must be carefully tuned. We implement a system around this model-based format-transforming encryption pipeline, and give an empirical analysis of its performance and (heuristic) security.

Data augmentation techniques, such as image transformations and combinations, are highly effective at improving the generalization of computer vision models, especially when training data is limited. However, such techniques are fundamentally incompatible with differentially private learning approaches, due to the latter’s built-in assumption that each training image’s contribution to the learned model is bounded. In this paper, we investigate why naive applications of multi-sample data augmentation techniques, such as mixup, fail to achieve good performance and propose two novel data augmentation techniques specifically designed for the constraints of differentially private learning. Our first technique, DP-Mix_Self, achieves SoTA classification performance across a range of datasets and settings by performing mixup on self-augmented data. Our second technique, DP-Mix_Diff, further improves performance by incorporating synthetic data from a pre-trained diffusion model into the mixup process. We open-source the code at https://github.com/wenxuan-Bao/DP-Mix.

Autonomous Vehicles (AVs) using LiDAR-based object detection systems are rapidly improving and becoming an increasingly viable method of transportation. While effective at perceiving the surrounding environment, these detection systems are shown to be vulnerable to attacks using lasers which can cause obstacle misclassifications or removal. These laser attacks, however, are challenging to perform, requiring precise aiming and accuracy. Our research exposes a new threat in the form of Intentional Electro-Magnetic-Interference (IEMI), which affects the time-of-flight (TOF) circuits that make up modern LiDARs. We show that these vulnerabilities can be exploited to force the AV Perception system to misdetect, misclassify objects, and perceive non-existent obstacles. We evaluate the vulnerability in three AV perception modules (PointPillars, PointRCNN, and Apollo) and show how the classification rate drops below 50%. We also analyze the impact of the IEMI injection on two fusion models (AVOD and Frustum-ConvNet) and in real-world scenarios. Finally, we discuss potential countermeasures and propose two strategies to detect signal injection.

How much does a machine learning algorithm leak about its training data, and why? Membership inference attacks are used as an auditing tool to quantify this leakage. In this paper, we present a comprehensivehypothesis testing framework that enables us not only to formally express the prior work in a consistent way, but also to design new membership inference attacks that use reference models to achieve a significantly higher power (true positive rate) for any (false positive rate) error. More importantly, we explainwhy different attacks perform differently. We present a template for indistinguishability games, and provide an interpretation of attack success rate across different instances of the game. We discuss various uncertainties of attackers that arise from the formulation of the problem, and show how our approach tries to minimize the attack uncertainty to the one bit secret about the presence or absence of a data point in the training set. We perform adifferential analysis between all types of attacks, explain the gap between them, and show what causes data points to be vulnerable to an attack (as the reasons vary due to different granularities of memorization, from overfitting to conditional memorization). Our auditing framework is openly accessible as part of thePrivacy Meter software tool.

Stalkerware is a form of malware that allows for the abusive monitoring of intimate partners. Primarily deployed on information-rich mobile platforms, these malicious applications allow for collecting information about a victim’s actions and behaviors, including location data, call audio, text messages, photos, and other personal details. While stalkerware has received increased attention from the security community, the ways in which stalkerware authors monetize their efforts have not been explored in depth. This paper represents the first large-scale technical analysis of monetization within the stalkerware ecosystem. We analyze the code base of 6,432 applications collected by the Coalition Against Stalkerware to determine their monetization strategies. We find that while far fewer stalkerware apps use ad libraries than normal apps, 99% of those that do use Google AdMob. We also find that payment services range from traditional in-app billing to cryptocurrency. Finally, we demonstrate that Google’s recent change to their Terms of Service (ToS) did not eliminate these applications, but instead caused a shift to other payment processors, while the apps can still be found on the Play Store; we verify through emulation that these apps often operate in blatant contravention of the ToS. Through this analysis, we find that the heterogeneity of markets and payment processors means that while point solutions can have impact on monetization, a multi-pronged solution involving multiple stakeholders is necessary to mitigate the financial incentive for developing stalkerware.

Differential privacy (DP) is a widely used notion for reasoning about privacy when publishing aggregate data. In this paper, we observe that certain DP mechanisms are amenable to a posteriori privacy analysis that exploits the fact that some outputs leak less information about the input database than others. To exploit this phenomenon, we introduce output differential privacy (ODP) and a new composition experiment, and leverage these new constructs to obtain significant privacy budget savings and improved privacy–utility tradeoffs under composition. All of this comes at no cost in terms of privacy; we do not weaken the privacy guarantee. To demonstrate the applicability of our a posteriori privacy analysis techniques, we analyze two well-known mechanisms: the Sparse Vector Technique and the Propose-Test-Release framework. We then show how our techniques can be used to save privacy budget in more general contexts: when a differentially private iterative mechanism terminates before its maximal number of iterations is reached, and when the output of a DP mechanism provides unsatisfactory utility. Examples of the former include iterative optimization algorithms, whereas examples of the latter include training a machine learning model with a large generalization error. Our techniques can be applied beyond the current paper to refine the analysis of existing DP mechanisms or guide the design of future mechanisms.

System-on-chip (SoC) developers utilize intellectual property (IP) cores from third-party vendors due to increasing design complexity, cost, as well as time-to-market constraints. A typical SoC consists of a wide variety of IP cores [such as processor, memory, controller, and field-programmable gate array (FPGA)] that interact using a network-on-chip (NoC). This global trend of designing SoCs using third-party IPs raises serious concerns about security vulnerabilities. Since NoC facilitates communication between all IPs in an SoC, NoC is the ideal place for any hardware Trojans to hide and launch a plethora of attacks. Due to the resource-constrained nature of SoCs, developing security solutions against such attacks is a major challenge. In particular, in an eavesdropping attack, a Trojan-infected router copies packets transferred through the NoC and reroutes the duplicated packets to an accompanying malicious application running on another IP in an attempt to extract confidential information. While authenticated encryption can thwart such attacks, it incurs unacceptable overhead in resource-constrained SoCs. In this article, we propose a lightweight alternative defense based on digital watermarking techniques. We develop theoretical models to provide security guarantees. Experiments using realistic SoC models and diverse applications demonstrate that our approach can significantly outperform state-of-the-art methods.

Automatic Speech Recognition (ASR) systems convert speech into text and can be placed into two broad categories: traditional and fully end-to-end. Both types have been shown to be vulnerable to adversarial audio examples that sound benign to the human ear but force the ASR to produce malicious transcriptions. Of these attacks, only the “psychoacoustic” attacks can create examples with relatively imperceptible perturbations, as they leverage the knowledge of the human auditory system. Unfortunately, existing psychoacoustic attacks can only be applied against traditional models, and are obsolete against the newer, fully end-to-end ASRs. In this paper, we propose an equalization-based psychoacoustic attack that can exploit both traditional and fully end-to-end ASRs. We successfully demonstrate our attack against real-world ASRs that include DeepSpeech and Wav2Letter. Moreover, we employ a user study to verify that our method creates low audible distortion. Specifically, 80 of the 100 participants voted in favor of \textit{all} our attack audio samples as less noisier than the existing state-of-the-art attack. Through this, we demonstrate both types of existing ASR pipelines can be exploited with minimum degradation to attack audio quality.

Speech and speaker recognition systems are employed in a variety of applications, from personal assistants to telephony surveillance and biometric authentication. The wide deployment of these systems has been made possible by the improved accuracy in neural networks. Like other systems based on neural networks, recent research has demonstrated that speech and speaker recognition systems are vulnerable to attacks using manipulated inputs. However, as we demonstrate in this paper, the end-to-end architecture of speech and speaker systems and the nature of their inputs make attacks and defenses against them substantially different than those in the image space. We demonstrate this first by systematizing existing research in this space and providing a taxonomy through which the community can evaluate future work. We then demonstrate experimentally that attacks against these models almost universally fail to transfer. In so doing, we argue that substantial additional work is required to provide adequate mitigations in this space.

We consider the problem of membership inference attacks on aggregate survey data through the use of several real-world datasets and a published study as a model for the survey. We apply membership inference attacks from the literature, and discover that methodological assumptions of existing attacks produce a misleading picture of the risk. When using a more realistic methodology, experiments reveal a more nuanced picture of the risk: membership inferences do succeed, but only a small subset of individuals are highly vulnerable to them. In fact, if the adversary wishes to avoid a high false positive rate, she should perform membership inferences only when she has a reason to believe that the target participated in the survey. However, our results do not imply that publishing survey data is inherently safe. Indeed, when applying membership inference not to individuals but to hospitals, we find that highly accurate inferences are possible.

Membership Inference Attacks (MIAs) aim to determine the presence of a record in a machine learning model's training data by querying the model. Recent work has demonstrated the effectiveness of MIA on various machine learning models and corresponding defenses have been proposed. However, both attacks and defenses have focused on an adversary that indiscriminately attacks all the records without regard to the cost of false positives or negatives. In this work, we revisit membership inference attacks from the perspective of a pragmatic adversary who carefully selects targets and make predictions conservatively. We design a new evaluation methodology that allows us to evaluate the membership privacy risk at the level of individuals and not only in aggregate. We experimentally demonstrate that highly vulnerable records exist even when the aggregate attack precision is close to 50% (baseline). Specifically, on the MNIST dataset, our pragmatic adversary achieves a precision of 95.05% whereas the prior attack only achieves a precision of 51.7%.

To protect database confidentiality even in the face of full compromise while supporting standard functionality, recent academic proposals and commercial products rely on a mix of encryption schemes. The recommendation is to apply strong, semantically secure encryption to the "sensitive" columns and protect other columns with property-revealing encryption (PRE) that supports operations such as sorting. We design, implement, and evaluate a new methodology for inferring data stored in such encrypted databases. The cornerstone is the multinomial attack , a new inference technique that is analytically optimal and empirically outperforms prior heuristic attacks against PRE-encrypted data. We also extend the multinomial attack to take advantage of correlations across multiple columns. This recovers PRE-encrypted data with sufficient accuracy to then apply machine learning and record linkage methods to infer columns protected by semantically secure encryption or redaction. We evaluate our methodology on medical, census, and union-membership datasets, showing for the first time how to infer full database records. For PRE-encrypted attributes such as demographics and ZIP codes, our attack outperforms the best prior heuristic by a factor of 16. Unlike any prior technique, we also infer attributes, such as incomes and medical diagnoses, protected by strong encryption. For example, when we infer that a patient in a hospital-discharge dataset has a mental health or substance abuse condition, this prediction is 97% accurate.

Side-channel risks of Intel's SGX have recently attracted great attention. Under the spotlight is the newly discovered page-fault attack, in which an OS-level adversary induces page faults to observe the page-level access patterns of a protected process running in an SGX enclave. With almost all proposed defense focusing on this attack, little is known about whether such efforts indeed raises the bar for the adversary, whether a simple variation of the attack renders all protection ineffective, not to mention an in-depth understanding of other attack surfaces in the SGX system. In the paper, we report the first step toward systematic analyses of side-channel threats that SGX faces, focusing on the risks associated with its memory management. Our research identifies 8 potential attack vectors, ranging from TLB to DRAM modules. More importantly, we highlight the common misunderstandings about SGX memory side channels, demonstrating that high frequent AEXs can be avoided when recovering EdDSA secret key through a new page channel and fine-grained monitoring of enclave programs (at the level of 64B) can be done through combining both cache and cross-enclave DRAM channels. Our findings reveal the gap between the ongoing security research on SGX and its side-channel weaknesses, redefine the side-channel threat model for secure enclaves, and can provoke a discussion on when to use such a system and how to use it securely.

We consider the problem of privacy-preserving data aggregation in a star network topology, i.e., several untrusting participants connected to a single aggregator. We require that the participants do not discover each other's data, and the service provider remains oblivious to each participant's individual contribution. Furthermore, the final result is to be published in a differentially private manner, i.e., the result should not reveal the contribution of any single participant to a (possibly external) adversary who knows the contributions of all other participants. In other words, we require a secure multiparty computation protocol that also incorporates a differentially private mechanism. Previous solutions have resorted to caveats such as postulating a trusted dealer to distribute keys to the participants, or introducing additional entities to withhold the decryption key from the aggregator, or relaxing the star topology by allowing pairwise communication amongst the participants. In this paper, we show how to obtain a noisy (differentially private) aggregation result using Shamir secret sharing and additively homomorphic encryption without these mitigating assumptions. More importantly, while we assume semi-honest participants, we allow the aggregator to be stronger than semi-honest, specifically in the sense that he can try to reduce the noise in the differentially private result.

Order-preserving encryption and its generalization order-revealing encryption (OPE/ORE) allow sorting, performing range queries, and filtering data - all while only having access to ciphertexts. But OPE and ORE ciphertexts necessarily leak information about plaintexts, and what level of security they provide in practice has been unclear. In this work, we introduce new leakage-abuse attacks that recover plaintexts from OPE/ORE-encrypted databases. Underlying our new attacks is a framework in which we cast the adversary's challenge as a non-crossing bipartite matching problem. This allows easy tailoring of attacks to a specific scheme's leakage profile. In a case study of customer records, we show attacks that recover 99% of first names, 97% of last names, and 90% of birthdates held in a database, despite all values being encrypted with the OPE scheme most widely used in practice. We also show the first attack against the recent frequency-hiding Kerschbaum scheme, to which no prior attacks have been demonstrated. Our attack recovers frequently occurring plaintexts most of the time.

Full-text search systems, such as Elasticsearch and Apache Solr, enable document retrieval based on keyword queries. In many deployments these systems are multi-tenant, meaning distinct users' documents reside in, and their queries are answered by, one or more shared search indexes. Large deployments may use hundreds of indexes across which user documents are randomly assigned. The results of a search query are filtered to remove documents to which a client should not have access. We show the existence of exploitable side channels in modern multi-tenant search. The starting point for our attacks is a decade-old observation that the TF-IDF scores used to rank search results can potentially leak information about other users' documents. To the best of our knowledge, no attacks have been shown that exploit this side channel in practice, and constructing a working side channel requires overcoming numerous challenges in real deployments. We nevertheless develop a new attack, called STRESS (Search Text RElevance Score Side channel), and in so doing show how an attacker can map out the number of indexes used by a service, obtain placement of a document within each index, and then exploit co-tenancy with all other users to (1) discover the terms in other tenants' documents or (2) determine the number of documents (belonging to other tenants) that contain a term of interest. In controlled experiments, we demonstrate the attacks on popular services such as GitHub and Xen.do. We conclude with a discussion of countermeasures.

Releasing full data records is one of the most challenging problems in data privacy. On the one hand, many of the popular techniques such as data de-identification are problematic because of their dependence on the background knowledge of adversaries. On the other hand, rigorous methods such as the exponential mechanism for differential privacy are often computationally impractical to use for releasing high dimensional data or cannot preserve high utility of original data due to their extensive data perturbation. This paper presents a criterion called plausible deniability that provides a formal privacy guarantee, notably for releasing sensitive datasets: an output record can be released only if a certain amount of input records are indistinguishable, up to a privacy parameter. This notion does not depend on the background knowledge of an adversary. Also, it can efficiently be checked by privacy tests. We present mechanisms to generate synthetic datasets with similar statistical properties to the input data and the same format. We study this technique both theoretically and experimentally. A key theoretical result shows that, with proper randomization, the plausible deniability mechanism generates differentially private synthetic data. We demonstrate the efficiency of this generative technique on a large dataset; it is shown to preserve the utility of original data with respect to various statistical analysis and machine learning measures.

Camouflaging user's actual location with fakes is a prevalent obfuscation technique for protecting location privacy. We show that the protection mechanisms based on the existing (ad hoc) techniques for generating fake locations are easily broken by inference attacks. They are also detrimental to many utility functions, as they fail to credibly imitate the mobility of living people. This paper introduces a systematic approach to synthesizing plausible location traces. We propose metrics that capture both geographic and semantic features of real location traces. Based on these statistical metrics, we design a privacy-preserving generative model to synthesize location traces which are plausible to be trajectories of some individuals with consistent lifestyles and meaningful mobilities. Using a state-of-the-art quantitative framework, we show that our synthetic traces can significantly paralyze location inference attacks. We also show that these fake traces have many useful statistical features in common with real traces, thus can be used in many geo-data analysis tasks. We guarantee that the process of generating synthetic traces itself is privacy preserving and ensures plausible deniability. Thus, although the crafted traces statistically resemble human mobility, they do not leak significant information about any particular individual whose data is used in the synthesis process.

To understand the gap between theory and practice for oblivious cloud storage, we experimentally evaluate four representative Oblivious RAM (ORAM) designs on Amazon S3. We replay realistic application traces to these ORAMs in order to understand whether they can meet the demands of various real applications using cloud storage as a backend. We find that metrics traditionally used in the ORAM literature, e.g., bandwidth overhead, fail to capture the practical needs of those applications. With a new understanding of the desirable properties, relevant metrics, and observations about the cloud services and their applications, we propose CURIOUS, a new modular partition-based ORAM framework, and show experimentally that it is thus far the most promising approach.

This paper presents a concrete and widespread example of situation where a user’s location privacy is unintentionally compromised by others, specifically the location-privacy threat that exists at access points (public hotspots, FON, home routers, etc.) that have a single public IP and make use of network address translation (NAT). As users connected to the same hotspot share a unique public IP address, a single user’s making a location-based request is enough to enable a service provider to map the IP address of the hotspot to its geographic coordinates, thus compromising the location privacy of all the other connected users. When successful, the service provider can locate users within a few hundreds of meters, thus improving over existing IP-location databases. Even in the case where IPs change periodically (e.g., by using DHCP), the service provider is still able to update a previous (IP, Location) mapping by inferring IP changes from authenticated communications (e.g., cookies). The contribution of this paper is three-fold: (i) We identify a novel location-privacy threat caused by shared public IPs in combination with NAT. (ii) We formalize and analyze the threat theoretically. In particular we derive and provide expressions of the probability that the service provider will learn the mapping and of the expected proportion of victims. (iii) We experimentally assess the state in practice by using real traces (collected from deployed hotspots over a period of 23 days) of users who accessed Google services. We also discuss how existing countermeasures can thwart the threat.