USENIX Security '22 Fall Accepted Papers

With the worldwide COVID-19 pandemic in 2020 and 2021 necessitating working from home, corporate Virtual Private Networks (VPNs) have become an important item securing the continued operation of companies around the globe. However, due to their different use case, corporate VPNs and how users interact with them differ from public VPNs, which are now commonly used by end-users.

In this paper, we present a first explorative study of eleven experts' and seven non-experts' mental models in the context of corporate VPNs. We find a partial alignment of these models in the high-level technical understanding while diverging in important parameters of how, when, and why VPNs are being used. While, in general, experts have a deeper technical understanding of VPN technology, we also observe that even they sometimes hold false beliefs on security aspects of VPNs. In summary, we show that the mental models of corporate VPNs differ from those for related security technology, e.g., HTTPS.

Our findings allow us to draft recommendations for practitioners to encourage a secure use of VPN technology (through training interventions, better communication, and system design changes in terms of device management). Furthermore, we identify avenues for future research, e.g., into experts' knowledge and balancing privacy and security between system operators and users.

DomainKeys Identified Mail (DKIM) is an email authentication protocol to protect the integrity of email contents. It has been proposed and standardized for over a decade and adopted by Yahoo!, Google, and other leading email service providers. However, little has been done to understand the adoption rate and potential security issues of DKIM due to the challenges of measuring DKIM deployment at scale.

In this paper, we provide a large-scale and longitudinal measurement study on how well DKIM is deployed and managed. Our study was made possible by a broad collection of datasets, including 9.5 million DKIM records from passive DNS datasets over five years and 460 million DKIM signatures from real-world email headers. Moreover, we conduct an active measurement on Alexa Top 1 million domains. Our measurement results show that 28.1% of Alexa Top 1 million domains have enabled DKIM, of which 2.9% are misconfigured. We demonstrate that the issues of DKIM key management and DKIM signatures are prevalent in the real world, even for well-known email providers (e.g., Gmail and Mail.ru). We recommend the security community should pay more attention to the systemic problems of DKIM deployment and mitigate these issues from the perspective of protocol design.

Secure two-party neural network inference (2PC-NN) can offer privacy protection for both the client and the server and is a promising technique in the machine-learning-as-a-service setting. However, the large overhead of the current 2PC-NN inference systems is still being a headache, especially when applied to deep neural networks such as ResNet50. In this work, we present Cheetah, a new 2PC-NN inference system that is faster and more communication-efficient than state-of-the-arts. The main contributions of Cheetah are two-fold: the first part includes carefully designed homomorphic encryption-based protocols that can evaluate the linear layers (namely convolution, batch normalization, and fully-connection) without any expensive rotation operation. The second part includes several lean and communication-efficient primitives for the non-linear functions (e.g., ReLU and truncation). Using Cheetah, we present intensive benchmarks over several large-scale deep neural networks. Take ResNet50 for an example, an end-to-end execution of Cheetah under a WAN setting costs less than 2.5 minutes and 2.3 gigabytes of communication, which outperforms CrypTFlow2 (ACM CCS 2020) by about 5.6× and 12.9×, respectively.

Electronic monitoring is the use of technology to track individuals accused or convicted of a crime (or civil violation) as an "alternative to incarceration." Traditionally, this technology has been in the form of ankle monitors, but recently federal, state, and local entities around the U.S. are shifting to using smartphone applications for electronic monitoring. These applications (apps) purport to make the monitoring simpler and more convenient for both the community supervisor and the person being monitored. However, due to the multipurpose nature of smartphones in people's lives and the amount of sensitive information (e.g., sensor data) smartphones make available, this introduces new risks to people coerced to use these apps.

To understand what type of privacy-related and other risks might be introduced to people who use these applications, we conducted a privacy-oriented analysis of 16 Android apps used for electronic monitoring. We analyzed the apps first technically, with static and (limited) dynamic analysis techniques. We also analyzed user reviews in the Google Play Store to understand the experiences of the people using these apps, and also the privacy policies. We found that apps contain numerous trackers, the permissions requested by them vary widely (with the most common one being location), and the reviews indicate that people find the apps invasive and frequently dysfunctional. We end the paper by encouraging mobile app marketplaces to reconsider their role in the future of electronic monitoring apps, and computer security and privacy researchers to consider their potential role in auditing carceral technologies. We hope that this work will lead to more transparency in this obfuscated ecosystem.

Given stigma and threats surrounding being gay or transgender, LGBTQ+ folks often seek support and information on navigating identity and personal (digital and physical) safety. While prior research on digital security advice focused on a general population and general advice, our work focuses on queer security, safety, and privacy advice-seeking to determine population-specific needs and takeaways for broader advice research. We conducted qualitative semi-structured interviews with 14 queer participants diverse across race, age, gender, sexuality, and socioeconomic status. We find that participants turn to their trusted queer support groups for advice, since they often experienced similar threats. We also document reasons that participants sometimes reject advice, including that it would interfere with their material livelihood and their potential to connect with others. Given our results, we recommend that queer-specific and general security and safety advice focus on specificity—why and how—over consistency, because advice cannot be one-size-fits-all. We also discuss the value of intersectionality as a framework for understanding vulnerability to harms in security research, since our participants' overlapping identities affected their threat models and advice perception.

The adversarial patch attack against image classification models aims to inject adversarially crafted pixels within a restricted image region (i.e., a patch) for inducing model misclassification. This attack can be realized in the physical world by printing and attaching the patch to the victim object; thus, it imposes a real-world threat to computer vision systems. To counter this threat, we design PatchCleanser as a certifiably robust defense against adversarial patches. In PatchCleanser, we perform two rounds of pixel masking on the input image to neutralize the effect of the adversarial patch. This image-space operation makes PatchCleanser compatible with any state-of-the-art image classifier for achieving high accuracy. Furthermore, we can prove that PatchCleanser will always predict the correct class labels on certain images against any adaptive white-box attacker within our threat model, achieving certified robustness. We extensively evaluate PatchCleanser on the ImageNet, ImageNette, and CIFAR-10 datasets and demonstrate that our defense achieves similar clean accuracy as state-of-the-art classification models and also significantly improves certified robustness from prior works. Remarkably, PatchCleanser achieves 83.9% top-1 clean accuracy and 62.1% top-1 certified robust accuracy against a 2%-pixel square patch anywhere on the image for the 1000-class ImageNet dataset.

Memory-safety issues in operating systems and popular applications are still top security threats. As one widely exploited vulnerability, Use After Free (UAF) resulted in hundreds of new incidents every year. Existing bug diagnosis techniques report the locations that allocate or deallocate the vulnerable object, but cannot provide sufficient information for developers to reason about a bug or synthesize a correct patch.

In this work, we identified incorrect reference counting as one common root cause of UAF bugs: if the developer forgets to increase the counter for a newly created reference, the program may prematurely free the actively used object, rendering other references dangling pointers. We call this problem reference miscounting. By proposing an omission-aware counting model, we developed an automatic method, FREEWILL, to diagnose UAF bugs. FREEWILL first reproduces a UAF bug and collects related execution trace. Then, it identifies the UAF object and related references. Finally, FREEWILL compares reference operations with our model to detect reference miscounting. We evaluated FREEWILL on 76 real-world UAF bugs and it successfully confirmed reference miscounting as root causes for 48 bugs and dangling usage for 18 bugs. FREEWILL also identified five null-pointer dereference bugs and failed to analyze five bugs. FREEWILL can complete its analysis within 15 minutes on average, showing its practicality for diagnosing UAF bugs.

As the performance of general-purpose processors faces diminishing improvements, computing systems are increasingly equipped with domain-specific accelerators. Today's high-end servers tightly integrate such accelerators with the CPU, e.g., giving them direct access to the CPU's last-level cache (LLC).

Caches are an important source of information leakage across security domains. This work explores combined cache attacks, complementing traditional co-tenancy with control over one or more accelerators. The constraints imposed on these accelerators, originally perceived as limitations, turn out to be advantageous to an attacker. We develop a novel approach for accelerators to find eviction sets, and leverage precise double-sided control over cache lines to expose undocumented behavior in non-inclusive Intel cache hierarchies.

We develop a compact and extensible FPGA hardware accelerator to demonstrate our findings. It constructs eviction sets at unprecedented speeds (<200 µs), outperforming existing techniques with one to three orders of magnitude. It maintains excellent performance, even under high noise pressure. We also use the accelerator to set up a covert channel with fine spatial granularity, encoding more than 3 bits per cache set. Furthermore, it can efficiently evict shared targets with tiny eviction sets, refuting the common assumption that eviction sets must be as large as the cache associativity.

A hypervisor is system software, managing and running virtual machines. Since the hypervisor is placed at the lowestlevel in the typical systems software stack, it has critical security implications. Once compromised, the entire software components running on top of the hypervisor (including all guest virtual machines and applications running within each guest virtual machine) are compromised as well, as the hypervisor has all the privileges to access those.

This paper proposes MUNDOFUZZ, a hypervisor fuzzer to enable both coverage-guided and grammar-aware fuzzing. We find that the coverage measurement in hypervisors suffers from noises due to the hypervisor's asynchronous system event handling. In order to filter out such noises, MUNDOFUZZ develops a statistical differential coverage measurement methods, allowing MUNDOFUZZ to capture the clean coverage information for hypervisor inputs. Moreover, we observe that hypervisor inputs have complex input grammars because it supports many different devices and each device has its own input format. Thus, MUNDOFUZZ learns the input grammar through inspecting the coverage characteristics of the given hypervisor input, which is based on the idea that the hypervisor behaves in a different way if the grammatically correct (or incorrect) input is given. We evaluated MUNDOFUZZ with popular hypervisors, QEMU and Bhyve, and MUNDOFUZZ outperformed other state-of-the-art hypervisor fuzzers ranging from 4.91% to 6.60% in terms of coverage. More importantly, MUNDOFUZZ identified 40 previously unknown bugs (including 9 CVEs), demonstrating its strong practical effectiveness in finding real-world hypervisor vulnerabilities.

Modern smartphones are equipped with Trusted Execution Environments (TEEs), offering security features resilient even against attackers able to fully compromise the normal operating system (e.g., Linux in Android devices). The academic community, as well as the smartphone manufacturers, have proposed to use TEEs to strengthen the security of authorization protocols. However, the usage of these protocols has been hampered by both practicality issues and lack of completeness in terms of security.

To address these issues, in this paper, we design, implement, and evaluate SARA (Secure Android Remote Authorization),an Android library that uses the existing TEE-powered Android APIs to implement secure, end-to-end remote authorization for Android apps. SARA is practical in its design, as it makes use of Android APIs and TEE features that are already present in modern Android devices to implement a novel secure authorization protocol. In fact, SARA does not require any modifications to the Android operating system nor to the code running in TrustZone (the TEE powering existing Android devices). For this reason, it can be readily used in existing apps running on existing smartphones. Moreover, SARA is designed to ensure that even developers that have no experience in implementing security protocols can make use of it within their apps. At the same time, SARA is secure, since it allows implementing authorization protocols that are resilient even against attackers able to achieve root privileges on a compromised Android device.

We first evaluate SARA by conducting a user study to ascertain its usability. Then, we prove SARA's security features by formally verifying its security protocol using ProVerif.

In this paper, we study the performance characteristics of nonbacktracking regex matchers and their vulnerability against ReDoS (regular expression denial of service) attacks. We focus on their known Achilles heel, which are extended regexes that use bounded quantifiers (e.g., '(ab){100}'). We propose a method for generating input texts that can cause ReDoS attacks on these matchers. The method exploits the bounded repetition and uses it to force expensive simulations of the deterministic automaton for the regex. We perform an extensive experimental evaluation of our and other state-of-the-art ReDoS generators on a large set of practical regexes with a comprehensive set of backtracking and nonbacktracking matchers, as well as experiments where we demonstrate ReDoS attacks on state-of-the-art real-world security applications containing SNORT with Hyperscan and the HW-accelerated regex matching engine on the NVIDIA BlueField-2 card. Our experiments show that bounded repetition is indeed a notable weakness of nonbacktracking matchers, with our generator being the only one capable of significantly increasing their running time.

Local Differential Privacy (LDP) protocols enable an untrusted server to perform privacy-preserving, federated data analytics. Various LDP protocols have been developed for different types of data such as categorical data, numerical data, and key-value data. Due to their distributed settings, LDP protocols are fundamentally vulnerable to poisoning attacks, in which fake users manipulate the server's analytics results via sending carefully crafted data to the server. However, existing poisoning attacks focused on LDP protocols for simple data types such as categorical and numerical data, leaving the security of LDP protocols for more advanced data types such as key-value data unexplored.

In this work, we aim to bridge the gap by introducing novel poisoning attacks to LDP protocols for key-value data. In such a LDP protocol, a server aims to simultaneously estimate the frequency and mean value of each key among some users, each of whom possesses a set of key-value pairs. Our poisoning attacks aim to simultaneously maximize the frequencies and mean values of some attacker-chosen target keys via sending carefully crafted data from some fake users to the sever. Specifically, since our attacks have two objectives, we formulate them as a two-objective optimization problem. Moreover, we propose a method to approximately solve the two-objective optimization problem, from which we obtain the optimal crafted data the fake users should send to the server. We demonstrate the effectiveness of our attacks to three LDP protocols for key-value data both theoretically and empirically. We also explore two defenses against our attacks, which are effective in some scenarios but have limited effectiveness in other scenarios. Our results highlight the needs for new defenses against our poisoning attacks.

Bridgefy is a messaging application that uses Bluetooth-based mesh networking. Its developers and others have advertised it for use in areas witnessing large-scale protests involving confrontations between protesters and state agents. In August 2020, a security analysis reported severe vulnerabilities that invalidated Bridgefy's claims of confidentiality, authentication, and resilience. In response, the developers adopted the Signal protocol and then continued to advertise their application as being suitable for use by higher-risk users.

In this work, we analyse the security of the revised Bridgefy messenger and SDK and invalidate its security claims. One attack (targeting the messenger) enables an adversary to compromise the confidentiality of private messages by exploiting a time-of-check to time-of-use (TOCTOU) issue, side-stepping Signal's guarantees. The other attack (targeting the SDK) allows an adversary to recover broadcast messages without knowing the network-wide shared encryption key.

We also found that the changes deployed in response to the August 2020 analysis failed to remedy the previously reported vulnerabilities. In particular, we show that (i) the protocol persisted to be susceptible to an active attacker-in-the-middle, (ii) an adversary continued to be able to impersonate other users in the broadcast channel of the Bridgefy messenger, (iii) the DoS attack using a decompression bomb was still applicable, albeit in a limited form, and that (iv) the privacy issues of Bridgefy remained largely unresolved.

As technology becomes an enabler of relationship abuse and coercive control, advocates who support survivors develop digital security practices to counter this. Existing research on technology-related abuse has primarily focused on describing the dynamics of abuse and developing solutions for this problem; we extend this literature by focusing on the security practices of advocates working "on the ground", i.e. in domestic violence shelters and other support services. We present findings from 26 semi-structured interviews and a data walkthrough workshop in which advocates described how they support survivors. We identified a variety of intertwined emotional and technical support practices, including establishing trust, safety planning, empowerment, demystification, supporting evidence collection and making referrals. By building relationships with other services and stakeholders, advocates also develop networks of care throughout society to create more supportive environments for survivors. Using critical and feminist theories, we see advocates as sources of crucial technical expertise to reduce this kind of violence in the future. Security and privacy researchers can build on and develop these networks of care by employing participatory methods and expanding threat modelling to account for interpersonal harms like coercive control and structural forms of discrimination such as misogyny and racism.

Fabricated media from deep learning models, or deepfakes, have been recently applied to facilitate social engineering efforts by constructing a trusted social persona. While existing works are primarily focused on deepfake detection, little is done to understand how users perceive and interact with deepfake persona (e.g., profiles) in a social engineering context. In this paper, we conduct a user study (n=286) to quantitatively evaluate how deepfake artifacts affect the perceived trustworthiness of a social media profile and the profile's likelihood to connect with users. Our study investigates artifacts isolated within a single media field (images or text) as well as mismatched relations between multiple fields. We also evaluate whether user prompting (or training) benefits users in this process. We find that artifacts and prompting significantly decrease the trustworthiness and request acceptance of deepfake profiles. Even so, users still appear vulnerable with 43% of them connecting to a deepfake profile under the best-case conditions. Through qualitative data, we find numerous reasons why this task is challenging for users, such as the difficulty of distinguishing text artifacts from honest mistakes and the social pressures entailed in the connection decisions. We conclude by discussing the implications of our results for content moderators, social media platforms, and future defenses.

Modern mobile games often contain in-app purchasing (IAP) for players to purchase digital items such as virtual currency, equipment, or extra moves. In theory, IAP should have been implemented securely; but in practice, we have found that many game developers have failed to do so, particularly by misplacing the trust of payment verification, e.g., by either locally verifying the payment transactions or without using any verification at all, leading to playing without paying vulnerabilities. This paper presents PAYMENTSCOPE, a static binary analysis tool to automatically identify vulnerable IAP implementations in mobile games. Through modeling of its IAP protocols with the SDK provided APIs using a payment-aware data flow analysis, PAYMENTSCOPE directly pinpoints untrusted payment verification vulnerabilities in game native binaries. We have implemented PAYMENTSCOPE on top of binary analysis framework Ghidra, and tested with 39,121 Unity (the most popular game engine) mobile games, with which PAYMENTSCOPE has identified 8,954 (22.89%) vulnerable games. Among them, 8,233 games do not verify the validity of payment transactions and 721 games simply verify the transactions locally. We have disclosed the identified vulnerabilities to developers of vulnerable games, and many of them have acknowledged our findings.

Internet censorship is widespread, impacting citizens of hundreds of countries around the world. Recent work has developed techniques that can perform widespread, longitudinal measurements of global Internet manipulation remotely and have focused largely on the scale of censorship measurements with minimal focus on reproducibility and consistency.

In this work we explore the role packet headers (e.g., source IP address and source port) have on DNS censorship. By performing a large-scale measurement study building on the techniques deployed by previous and current censorship measurement platforms, we find that choice of ephemeral source port and local source IP address (e.g., x.x.x.7 vs x.x.x.8) influence routing, which in turn influences DNS censorship. We show that 37% of IPs across 56% ASes measured show some change in censorship behavior depending on source port and local source IP. This behavior is frequently all-or-nothing, where choice of header can result in no observable censorship. Such behavior mimics and could be misattributed to geolocation error, packet loss, or network outages. The scale of censorship differences can more than double depending on the lowest 3 bits of the source IP address, consistent with known router load balancing techniques. We also observe smaller-scale censorship variation where only a few domains experience censorship differences based on packet parameters. We lastly find that these variations are persistent; packet retries do not control for observed variation. Our results point to the need for methodological changes in future DNS censorship measurement, which we discuss.

Quasi-identifier-based deidentification techniques (QI-deidentification) are widely used in practice, including k-anonymity, l-diversity, and t-closeness. We present three new attacks on QI-deidentification: two theoretical attacks and one practical attack on a real dataset. In contrast to prior work, our theoretical attacks work even if every attribute is a quasi-identifier. Hence, they apply to k-anonymity, l-diversity, t-closeness, and most other QI-deidentification techniques.

First, we introduce a new class of privacy attacks called downcoding attacks, and prove that every QI-deidentification scheme is vulnerable to downcoding attacks if it is minimal and hierarchical. Second, we convert the downcoding attacks into powerful predicate singling-out (PSO) attacks, which were recently proposed as a way to demonstrate that a privacy mechanism fails to legally anonymize under Europe's General Data Protection Regulation. Third, we use LinkedIn.com to reidentify 3 students in a k-anonymized dataset published by EdX (and show thousands are potentially vulnerable), undermining EdX's claimed compliance with the Family Educational Rights and Privacy Act.

The significance of this work is both scientific and political. Our theoretical attacks demonstrate that QI-deidentification may offer no protection even if every attribute is treated as a quasi-identifier. Our practical attack demonstrates that even deidentification experts acting in accordance with strict privacy regulations fail to prevent real-world reidentification. Together, they rebut a foundational tenet of QI-deidentification and challenge the actual arguments made to justify the continued use of k-anonymity and other QI-deidentification techniques.

Online tracking has garnered significant attention due to the privacy risk it poses to users. Among the various approaches, techniques that identify which extensions are installed in a browser can be used for fingerprinting browsers and tracking users, but also for inferring personal and sensitive user data. While preventing certain fingerprinting techniques is relatively simple, mitigating behavior-based extension-fingerprinting poses a significant challenge as it relies on hiding actions that stem from an extension's functionality. To that end, we introduce the concept of DOM Reality Shifting, whereby we split the reality users experience while browsing from the reality that webpages can observe. To demonstrate our approach we develop Simulacrum, a prototype extension that implements our defense through a targeted instrumentation of core Web API interfaces. Despite being conceptually straightforward, our implementation highlights the technical challenges posed by the complex and often idiosyncratic nature and behavior of web applications, modern browsers, and the JavaScript language. We experimentally evaluate our system against a state-of-theart DOM-based extension fingerprinting system and find that Simulacrum readily protects 95.37% of susceptible extensions. We then identify trivial modifications to extensions that enable our defense for the majority of the remaining extensions. To facilitate additional research and protect users from privacy-invasive behaviors we will open-source our system.

Despite the fact that most real-world software systems today are written in multiple programming languages, existing program analysis based security techniques are still limited to single-language code. In consequence, security flaws (e.g., code vulnerabilities) at and across language boundaries are largely left out as blind spots. We present PolyCruise, a technique that enables holistic dynamic information flow analysis (DIFA) across heterogeneous languages hence security applications empowered by DIFA (e.g., vulnerability discovery) for multilingual software. PolyCruise combines a light language-specific analysis that computes symbolic dependencies in each language unit with a language-agnostic online data flow analysis guided by those dependencies, in a way that overcomes language heterogeneity. Extensive evaluation of its implementation for Python-C programs against micro, medium-sized, and large-scale benchmarks demonstrated PolyCruise's practical scalability and promising capabilities. It has enabled the discovery of 14 unknown cross-language security vulnerabilities in real-world multilingual systems such as NumPy, with 11 confirmed, 8 CVEs assigned, and 8 fixed so far. We also contributed the first benchmark suite for systematically assessing multilingual DIFA.

With the recent report of erroneous content in 3GPP specifications leading to real-world vulnerabilities, attention has been drawn to not only the specifications but also the way they are maintained and adopted by manufacturers and carriers. In this paper, we report the first study on this 3GPP ecosystem, for the purpose of understanding its security hazards. Our research leverages 414,488 Change Requests (CRs) that document the problems discovered from specifications and proposed changes, which provides valuable information about the security assurance of the 3GPP ecosystem.

Analyzing these CRs is impeded by the challenge in finding security-relevant CRs (SR-CRs), whose security connections cannot be easily established by even human experts. To identify them, we developed a novel NLP/ML pipeline that utilizes a small set of positively labeled CRs to recover 1,270 high-confidence SR-CRs. Our measurement on them reveals serious consequences of specification errors and their causes, including design errors and presentation issues, particularly the pervasiveness of inconsistent descriptions (misalignment) in security-relevant content. Also important is the discovery of a security weakness inherent to the 3GPP ecosystem, which publishes an SR-CR long before the specification has been fixed and related systems have been patched. This opens an "attack window", which can be as long as 11 years! Interestingly, we found that some recently reported vulnerabilities are actually related to the CRs published years ago. Further, we identified a set of vulnerabilities affecting major carriers and mobile phones that have not been addressed even today. With the trend of SR-CRs not showing any sign of abating, we propose measures to improve the security assurance of the ecosystem, including responsible handling of SR-CRs.

Web users enter their email addresses into online forms for a variety of reasons, including signing in or signing up for a service or subscribing to a newsletter. While enabling such functionality, email addresses typed into forms can also be collected by third-party scripts even when users change their minds and leave the site without submitting the form. Email addresses—or identifiers derived from them—are known to be used by data brokers and advertisers for cross-site, cross-platform, and persistent identification of potentially unsuspecting individuals. In order to find out whether access to online forms is misused by online trackers, we present a measurement of email and password collection that occurs before the form submission on the top 100,000 websites. We evaluate the effect of user location, browser configuration, and interaction with consent dialogs by comparing results across two vantage points (EU/US), two browser configurations (desktop/mobile), and three consent modes. Our crawler finds and fills email and password fields, monitors the network traffic for leaks, and intercepts script access to filled input fields. Our analyses show that users' email addresses are exfiltrated to tracking, marketing and analytics domains before form submission and without giving consent on 1,844 websites in the EU crawl and 2,950 websites in the US crawl. While the majority of email addresses are sent to known tracking domains, we further identify 41 tracker domains that are not listed by any of the popular blocklists. Furthermore, we find incidental password collection on 52 websites by third-party session replay scripts.

To enable safe and reliable decision-making, autonomous vehicles (AVs) feed sensor data to perception algorithms to understand the environment. Sensor fusion with multi-frame tracking is becoming increasingly popular for detecting 3D objects. Thus, in this work, we perform an analysis of camera-LiDAR fusion, in the AV context, under LiDAR spoofing attacks. Recently, LiDAR-only perception was shown vulnerable to LiDAR spoofing attacks; however, we demonstrate these attacks are not capable of disrupting camera-LiDAR fusion. We then define a novel, context-aware attack: frustum attack, and show that out of 8 widely used perception algorithms – across 3 architectures of LiDAR-only and 3 architectures of camera-LiDAR fusion – all are significantly vulnerable to the frustum attack. In addition, we demonstrate that the frustum attack is stealthy to existing defenses against LiDAR spoofing as it preserves consistencies between camera and LiDAR semantics. Finally, we show that the frustum attack can be exercised consistently over time to form stealthy longitudinal attack sequences, compromising the tracking module and creating adverse outcomes on end-to-end AV control.

We present the first over-the-air attack on IEEE 802.15.4z High-Rate Pulse Repetition Frequency (HRP) Ultra-Wide Band (UWB) distance measurement systems. Specifically, we demonstrate a practical distance reduction attack against pairs of Apple U1 chips (embedded in iPhones and AirTags), as well as against U1 chips inter-operating with NXP and Qorvo UWB chips. These chips have been deployed in a wide range of phones and cars to secure car entry and start and are projected for secure contactless payments, home locks, and contact tracing systems. Our attack operates without any knowledge of cryptographic material, results in distance reductions from 12m (actual distance) to 0m (spoofed distance) with attack success probabilities of up to 4%, and requires only an inexpensive (USD 65) off-the-shelf device. Access control can only tolerate sub-second latencies to not inconvenience the user, leaving little margin to perform time-consuming verifications. These distance reductions bring into question the use of UWB HRP in security-critical applications.

Equality operators are an essential building block in tasks over secure computation such as private information retrieval. In private information retrieval (PIR), a user queries a database such that the server does not learn which element is queried. In this work, we propose equality operators for constant-weight codewords. A constant-weight code is a collection of codewords that share the same Hamming weight. Constant-weight equality operators have a multiplicative depth that depends only on the Hamming weight of the code, not the bit-length of the elements. In our experiments, we show how these equality operators are up to 10 times faster than existing equality operators. Furthermore, we propose PIR using the constant-weight equality operator or constant-weight PIR, which is a PIR protocol using an approach previously deemed impractical. We show that for private retrieval of large, streaming data, constant-weight PIR has a smaller communication complexity and lower runtime compared to SEALPIR and MulPIR, respectively, which are two state-of-the-art solutions for PIR. Moreover, we show how constant-weight PIR can be extended to keyword PIR. In keyword PIR, the desired element is retrieved by a unique identifier pertaining to the sought item, e.g., the name of a file. Previous solutions to keyword PIR require one or multiple rounds of communication to reduce the problem to normal PIR. We show that constant-weight PIR is the first practical single-round solution to single-server keyword PIR.

Serverless computing has freed developers from the burden of managing their own platform and infrastructure, allowing them to rapidly prototype and deploy applications. Despite its surging popularity, however, serverless raises a number of concerning security implications. Among them is the difficulty of investigating intrusions – by decomposing traditional applications into ephemeral re-entrant functions, serverless has enabled attackers to conceal their activities within legitimate workflows, and even prevent root cause analysis by abusing warm container reuse policies to break causal paths. Unfortunately, neither traditional approaches to system auditing nor commercial serverless security products provide the transparency needed to accurately track these novel threats.

In this work, we propose ALASTOR, a provenance-based auditing framework that enables precise tracing of suspicious events in serverless applications. ALASTOR records function activity at both system and application layers to capture a holistic picture of each function instances' behavior. It then aggregates provenance from different functions at a central repository within the serverless platform, stitching it together to produce a global data provenance graph of complex function workflows. ALASTOR is both function and language-agnostic, and can easily be integrated into existing serverless platforms with minimal modification. We implement ALASTOR for the OpenFaaS platform and evaluate its performance using the well-established Nordstrom Hello,Retail! application, discovering in the process that ALASTOR imposes manageable overheads (13.74%), in exchange for significantly improved forensic capabilities as compared to commercially-available monitoring tools. To our knowledge, ALASTOR is the first auditing framework specifically designed to satisfy the operational requirements of serverless platforms.

Machine unlearning, i.e. having a model forget about some of its training data, has become increasingly more important as privacy legislation promotes variants of the right-to-be-forgotten. In the context of deep learning, approaches for machine unlearning are broadly categorized into two classes: exact unlearning methods, where an entity has formally removed the data point's impact on the model by retraining the model from scratch, and approximate unlearning, where an entity approximates the model parameters one would obtain by exact unlearning to save on compute costs. In this paper, we first show that the definition that underlies approximate unlearning, which seeks to prove the approximately unlearned model is close to an exactly retrained model, is incorrect because one can obtain the same model using different datasets. Thus one could unlearn without modifying the model at all. We then turn to exact unlearning approaches and ask how to verify their claims of unlearning. Our results show that even for a given training trajectory one cannot formally prove the absence of certain data points used during training. We thus conclude that unlearning is only well-defined at the algorithmic level, where an entity's only possible auditable claim to unlearning is that they used a particular algorithm designed to allow for external scrutiny during an audit.

Membership inference attacks are a key measure to evaluate privacy leakage in machine learning (ML) models. It is important to train ML models that have high membership privacy while largely preserving their utility. In this work, we propose a new framework to train privacy-preserving models that induce similar behavior on member and non-member inputs to mitigate membership inference attacks. Our framework, called SELENA, has two major components. The first component and the core of our defense is a novel ensemble architecture for training. This architecture, which we call Split-AI, splits the training data into random subsets, and trains a model on each subset of the data. We use an adaptive inference strategy at test time: our ensemble architecture aggregates the outputs of only those models that did not contain the input sample in their training data. Our second component, Self-Distillation, (self-)distills the training dataset through our Split-AI ensemble, without using any external public datasets. We prove that our Split-AI architecture defends against a family of membership inference attacks, however, our defense does not provide provable guarantees against all possible attackers as opposed to differential privacy. This enables us to improve the utility of models compared to DP. Through extensive experiments on major benchmark datasets we show that SELENA presents a superior trade-off between (empirical) membership privacy and utility compared to the state of the art empirical privacy defenses. In particular, SELENA incurs no more than 3.9% drop in classification accuracy compared to the undefended model while reducing the membership inference attack advantage by a factor of up to 3.7 compared to MemGuard and a factor of up to 2.1 compared to adversarial regularization.

This paper introduces Otti, a general-purpose compiler for (zk)SNARKs that provides support for numerical optimization problems. Otti produces efficient arithmetizations of programs that contain optimization problems including linear programming (LP), semi-definite programming (SDP), and a broad class of stochastic gradient descent (SGD) instances. Numerical optimization is a fundamental algorithmic building block: applications include scheduling and resource allocation tasks, approximations to NP-hard problems, and training of neural networks. Otti takes as input arbitrary programs written in a subset of C that contain optimization problems specified via an easy-to-use API. Otti then automatically produces rank-1 constraint satisfiability (R1CS) instances that express a succinct transformation of those programs. Correct execution of the transformed program implies the optimality of the solution to the original optimization problem. Our evaluation on real benchmarks shows that Otti, instantiated with the Spartan proof system, can prove the optimality of solutions in zero-knowledge in as little as 100 ms—over 4 orders of magnitude faster than existing approaches.

Neural network pruning has been an essential technique to reduce the computation and memory requirements for using deep neural networks for resource-constrained devices. Most existing research focuses primarily on balancing the sparsity and accuracy of a pruned neural network by strategically removing insignificant parameters and retraining the pruned model. Such efforts on reusing training samples pose serious privacy risks due to increased memorization, which, however, has not been investigated yet.

In this paper, we conduct the first analysis of privacy risks in neural network pruning. Specifically, we investigate the impacts of neural network pruning on training data privacy, i.e., membership inference attacks. We first explore the impact of neural network pruning on prediction divergence, where the pruning process disproportionately affects the pruned model's behavior for members and non-members. Meanwhile, the influence of divergence even varies among different classes in a fine-grained manner. Enlightened by such divergence, we proposed a self-attention membership inference attack against the pruned neural networks. Extensive experiments are conducted to rigorously evaluate the privacy impacts of different pruning approaches, sparsity levels, and adversary knowledge. The proposed attack shows the higher attack performance on the pruned models when compared with eight existing membership inference attacks. In addition, we propose a new defense mechanism to protect the pruning process by mitigating the prediction divergence based on KL-divergence distance, whose effectiveness has been experimentally demonstrated to effectively mitigate the privacy risks while maintaining the sparsity and accuracy of the pruned models.

VPN adoption has seen steady growth over the past decade due to increased public awareness of privacy and surveillance threats. In response, certain governments are attempting to restrict VPN access by identifying connections using "dual use" DPI technology. To investigate the potential for VPN blocking, we develop mechanisms for accurately fingerprinting connections using OpenVPN, the most popular protocol for commercial VPN services. We identify three fingerprints based on protocol features such as byte pattern, packet size, and server response. Playing the role of an attacker who controls the network, we design a two-phase framework that performs passive fingerprinting and active probing in sequence. We evaluate our framework in partnership with a million-user ISP and find that we identify over 85% of OpenVPN flows with only negligible false positives, suggesting that OpenVPN-based services can be effectively blocked with little collateral damage. Although some commercial VPNs implement countermeasures to avoid detection, our framework successfully identified connections to 34 out of 41 "obfuscated" VPN configurations. We discuss the implications of the VPN fingerprintability for different threat models and propose short-term defenses. In the longer term, we urge commercial VPN providers to be more transparent about their obfuscation approaches and to adopt more principled detection countermeasures, such as those developed in censorship circumvention research.

The widespread availability of vulnerable IoT devices has resulted in IoT botnets. A particularly concerning IoT botnet can be built around high-wattage IoT devices such as EV chargers because, in large numbers, they can abruptly change the electricity consumption in the power grid. These attacks are called Manipulation of Demand via IoT (MaDIoT) attacks. Previous research has shown that the existing power grid protection mechanisms prevent any large-scale negative consequences to the grid from MaDIoT attacks. In this paper, we analyze this assumption and show that an intelligent attacker with extra knowledge about the power grid and its state, can launch more sophisticated attacks. Rather than attacking all locations at random times, our adversary uses an instability metric that lets the attacker know the specific time and geographical location to activate the high-wattage bots. We call these new attacks MaDIoT 2.0.

Generative machine learning models have made convincing voice synthesis a reality. While such tools can be extremely useful in applications where people consent to their voices being cloned (e.g., patients losing the ability to speak, actors not wanting to have to redo dialog, etc), they also allow for the creation of nonconsensual content known as deepfakes. This malicious audio is problematic not only because it can convincingly be used to impersonate arbitrary users, but because detecting deepfakes is challenging and generally requires knowledge of the specific deepfake generator. In this paper, we develop a new mechanism for detecting audio deepfakes using techniques from the field of articulatory phonetics. Specifically, we apply fluid dynamics to estimate the arrangement of the human vocal tract during speech generation and show that deepfakes often model impossible or highly-unlikely anatomical arrangements. When parameterized to achieve 99.9% precision, our detection mechanism achieves a recall of 99.5%, correctly identifying all but one deepfake sample in our dataset. We then discuss the limitations of this approach, and how deepfake models fail to reproduce all aspects of speech equally. In so doing, we demonstrate that subtle, but biologically constrained aspects of how humans generate speech are not captured by current models, and can therefore act as a powerful tool to detect audio deepfakes.

Network side channels (NSCs) leak secrets through packet timing and packet sizes. They are of particular concern in public IaaS Clouds, where any tenant may be able to colocate and indirectly observe a victim's traffic shape. We present Pacer, the first system that eliminates NSC leaks in public IaaS Clouds end-to-end. It builds on the principled technique of shaping guest traffic outside the guest to make the traffic shape independent of secrets by design. However, Pacer also addresses important concerns that have not been considered in prior work—it prevents internal side-channel leaks from affecting reshaped traffic, and it respects network flow control, congestion control and loss recovery signals. Pacer is implemented as a paravirtualizing extension to the host hypervisor, requiring modest changes to the hypervisor and the guest kernel and optional, minimal changes to applications. We present Pacer's key abstraction of a cloaked tunnel, describe its design and implementation, and show through an experimental evaluation that Pacer imposes moderate overheads on bandwidth, client latency, and server throughput, while thwarting attacks using state-of-the-art CNN classifiers.

The increasing complexity of modern processors poses many challenges to existing hardware verification tools and methodologies for detecting security-critical bugs. Recent attacks on processors have shown the fatal consequences of uncovering and exploiting hardware vulnerabilities.

Fuzzing has emerged as a promising technique for detecting software vulnerabilities. Recently, a few hardware fuzzing techniques have been proposed. However, they suffer from several limitations, including non-applicability to commonly used hardware description languages (HDLs) like Verilog and VHDL, the need for significant human intervention, and inability to capture many intrinsic hardware behaviors, such as signal transitions and floating wires.

In this paper, we present the design and implementation of a novel hardware fuzzer, TheHuzz, that overcomes the aforementioned limitations and significantly improves the state of the art. We analyze the intrinsic behaviors of hardware designs in HDLs and then measure the coverage metrics that model such behaviors. TheHuzz generates assembly-level instructions to increase the desired coverage values, thereby finding many hardware bugs that are exploitable from software. We evaluate TheHuzz on four popular open-source processors and achieve 1.98× and 3.33× the speed compared to the industry-standard random regression approach and the state-of-the-art hardware fuzzer, DifuzzRTL, respectively. Using TheHuzz, we detected 11 bugs in these processors, including 8 new bugs, and we demonstrate exploits using the detected bugs. We also show that TheHuzz overcomes the limitations of formal verification tools from the semiconductor industry by comparing its findings to those discovered by the Cadence JasperGold tool.

Branch Target Injection (BTI or Spectre v2) is one of the most dangerous transient execution vulnerabilities, as it allows an attacker to abuse indirect branch mispredictions to leak sensitive information. Unfortunately, it also has proven difficult to mitigate, with vendors originally resorting to inefficient software mitigations like retpoline. Recently, efficient hardware mitigations such as Intel eIBRS and Arm CSV2 have been deployed as a replacement in production, isolating the branch target state across privilege domains. The assumption is that this is sufficient to deter practical BTI exploitation. In this paper, we challenge this belief and disclose fundamental design flaws in both Intel and Arm solutions.

We introduce Branch History Injection (BHI or Spectre-BHB), a new primitive to build cross-privilege BTI attacks on systems deploying isolation-based hardware defenses. BHI builds on the observation that, while the branch target state is now isolated across privilege domains, such isolation is not extended to other branch predictor elements tracking the branch history state—ultimately re-enabling cross-privilege attacks. We further analyze the guarantees of a hypothetical isolation-based mitigation which also isolates the branch history and show that, barring a collision-free design, practical same-predictor-mode attacks are still possible. To instantiate our approach, we present end-to-end exploits leaking kernel memory from userland on Intel systems at 160 bytes/s, in spite of existing or hypothetical isolation-based mitigations. We conclude software defenses such as retpoline remain the only practical BTI mitigations in the foreseeable future and the pursuit for efficient hardware mitigations must continue.

Increasing use of machine learning (ML) technologies in privacy-sensitive domains such as medical diagnoses, lifestyle predictions, and business decisions highlights the need to better understand if these ML technologies are introducing leakage of sensitive and proprietary training data. In this paper, we focus on model inversion attacks where the adversary knows non-sensitive attributes about records in the training data and aims to infer the value of a sensitive attribute unknown to the adversary, using only black-box access to the target classification model. We first devise a novel confidence score-based model inversion attribute inference attack that significantly outperforms the state-of-the-art. We then introduce a label-only model inversion attack that relies only on the model's predicted labels but still matches our confidence score-based attack in terms of attack effectiveness. We also extend our attacks to the scenario where some of the other (non-sensitive) attributes of a target record are unknown to the adversary. We evaluate our attacks on two types of machine learning models, decision tree and deep neural network, trained on three real datasets. Moreover, we empirically demonstrate the disparate vulnerability of model inversion attacks, i.e., specific groups in the training dataset (grouped by gender, race, etc.) could be more vulnerable to model inversion attacks.

We address the challenging problem of efficient trust establishment in constrained networks, i.e., networks that are composed of a large and dynamic set of (possibly heterogeneous) devices with limited bandwidth, connectivity, storage, and computational capabilities. Constrained networks are an integral part of many emerging application domains, from IoT meshes to satellite networks. A particularly difficult challenge is how to enforce timely revocation of compromised or faulty devices. Unfortunately, current solutions and techniques cannot cope with idiosyncrasies of constrained networks, since they mandate frequent real-time communication with centralized entities, storage and maintenance of large amounts of revocation information, and incur considerable bandwidth overhead.

To address the shortcomings of existing solutions, we design V'CER, a secure and efficient scheme for certificate validation that augments and benefits a PKI for constrained networks. V'CER utilizes unique features of Sparse Merkle Trees (SMTs) to perform lightweight revocation checks, while enabling collaborative operations among devices to keep them up-to-date when connectivity to external authorities is limited. V'CER can complement any PKI scheme to increase its flexibility and applicability, while ensuring fast dissemination of validation information independent of the network routing or topology. V'CER requires under 3KB storage per node covering 10⁶ certificates. We developed and deployed a prototype of V'CER on an in-orbit satellite and our large-scale simulations demonstrate that V'CER decreases the number of requests for updates from external authorities by over 93%, when nodes are intermittently connected.

Mobile applications (apps) often delegate their own functions to other parties, which makes them become a super ecosystem hosting these parties. Therefore, such mobile apps are being called super-apps, and the delegated parties are subsequently called sub-apps, behaving like "app-in-app". Sub-apps not only load (third-party) resources like a normal app, but also have access to the privileged APIs provided by the super-app. This leads to an important research question—determining who can access these privileged APIs.

Real-world super-apps, according to our study, adopt three types of identities—namely web domains, sub-app IDs, and capabilities—to determine privileged API access. However, existing identity checks of these three types are often not well designed, leading to a disobey of the least privilege principle. That is, the granted recipient of a privileged API is broader than intended, thus defined as an "identity confusion" in this paper. To the best of our knowledge, no prior works have studied this type of identity confusion vulnerability.

In this paper, we perform the first systematic study of identity confusion in real-world app-in-app ecosystems. We find that confusions of the aforementioned three types of identities are widespread among all 47 studied super-apps. More importantly, such confusions lead to severe consequences such as manipulating users' financial accounts and installing malware on a smartphone. We responsibly reported all of our findings to developers of affected super-apps, and helped them to fix their vulnerabilities.

Federated Learning (FL) is a collaborative machine learning approach allowing participants to jointly train a model without having to share their private, potentially sensitive local datasets with others. Despite its benefits, FL is vulnerable to so-called backdoor attacks, in which an adversary injects manipulated model updates into the federated model aggregation process so that the resulting model will provide targeted false predictions for specific adversary-chosen inputs. Proposed defenses against backdoor attacks based on detecting and filtering out malicious model updates consider only very specific and limited attacker models, whereas defenses based on differential privacy-inspired noise injection significantly deteriorate the benign performance of the aggregated model. To address these deficiencies, we introduce FLAME, a defense framework that estimates the sufficient amount of noise to be injected to ensure the elimination of backdoors. To minimize the required amount of noise, FLAME uses a model clustering and weight clipping approach. This ensures that FLAME can maintain the benign performance of the aggregated model while effectively eliminating adversarial backdoors. Our evaluation of FLAME on several datasets stemming from application areas including image classification, word prediction, and IoT intrusion detection demonstrates that FLAME removes backdoors effectively with a negligible impact on the benign performance of the models.