USENIX Security '22 Winter Accepted Papers

Targeted deanonymization attacks let a malicious website discover whether a website visitor bears a certain public identifier, such as an email address or a Twitter handle. These attacks were previously considered to rely on several assumptions, limiting their practical impact. In this work, we challenge these assumptions and show the attack surface for deanonymization attacks is drastically larger than previously considered. We achieve this by using the cache side channel for our attack, instead of relying on cross-site leaks. This makes our attack oblivious to recently proposed software-based isolation mechanisms, including cross-origin resource policies (CORP), cross-origin opener policies (COOP) and SameSite cookie attribute. We evaluate our attacks on multiple hardware microarchitectures, multiple operating systems and multiple browser versions, including the highly-secure Tor Browser, and demonstrate practical targeted deanonymization attacks on major sites, including Google, Twitter, LinkedIn, TikTok, Facebook, Instagram and Reddit. Our attack runs in less than 3 seconds in most cases, and can be scaled to target an exponentially large amount of users.

To stop these attacks, we present a full-featured defense deployed as a browser extension. To minimize the risk to vulnerable individuals, our defense is already available on the Chrome and Firefox app stores. We have also responsibly disclosed our findings to multiple tech vendors, as well as to the Electronic Frontier Foundation. Finally, we provide guidance to websites and browser vendors, as well as to users who cannot install the extension.

Deep learning systems are known to be vulnerable to adversarial examples. In particular, query-based black-box attacks do not require knowledge of the deep learning model, but can compute adversarial examples over the network by submitting queries and inspecting returns. Recent work largely improves the efficiency of those attacks, demonstrating their practicality on today's ML-as-a-service platforms.

We propose Blacklight, a new defense against query-based black-box adversarial attacks. Blacklight is driven by a fundamental insight: to compute adversarial examples, these attacks perform iterative optimization over the network, producing queries highly similar in the input space. Thus Blacklight detects query-based black-box attacks by detecting highly similar queries, using an efficient similarity engine operating on probabilistic content fingerprints. We evaluate Blacklight against eight state-of-the-art attacks, across a variety of models and image classification tasks. Blacklight identifies them all, often after only a handful of queries. By rejecting all detected queries, Blacklight prevents any attack from completing, even when persistent attackers continue to submit queries after banned accounts or rejected queries. Blacklight is also robust against several powerful countermeasures, including an optimal black-box attack that approximates white-box attacks in efficiency. Finally, we illustrate how Blacklight generalizes to other domains like text classification.

Adversarial attacks can fool deep learning models by imposing imperceptible perturbations onto natural examples, which have provoked concerns in various security-sensitive applications. Among them, decision-based black-box attacks are practical yet more challenging, where the adversary can only acquire the final classification labels by querying the target model without access to the model's details. Under this setting, existing works usually rely on heuristics and exhibit unsatisfactory performance in terms of query efficiency and attack success rate. To better understand the rationality of these heuristics and further improve over existing methods, we propose AutoDA to automatically discover decision-based iterative adversarial attack algorithms. In our approach, we construct a generic search space of attack algorithms and develop an efficient search algorithm to explore this space. Although we adopt a small and fast model to efficiently evaluate and discover qualified attack algorithms during the search, extensive experiments demonstrate that the discovered algorithms are simple yet query-efficient when attacking larger models on the CIFAR-10 and ImageNet datasets. They achieve comparable performance with the human-designed state-of-the-art decision-based iterative attack methods consistently.

Studying developers is an important aspect of usable security and privacy research. In particular, studying security development challenges such as the usability of security APIs, the secure use of information sources during development or the effectiveness of IDE security plugins raised interest in recent years. However, recruiting skilled participants with software development experience is particularly challenging, and it is often not clear what security researchers can expect from certain participant samples, which can make research results hard to compare and interpret. Hence, in this work, we study for the first time opportunities and challenges of different platforms to recruit participants with software development experience for security development studies. First, we identify popular recruitment platforms in 59 papers. Then, we conduct a comparative online study with 706 participants based on self-reported software development experience across six recruitment platforms. Using an online questionnaire, we investigate participants' programming and security experiences, skills and knowledge. We find that participants across all samples report rich general software development and security experience, skills, and knowledge. Based on our results, we recommend developer recruitment from Upwork for practical coding studies and Amazon MTurk along with a pre-screening survey to reduce additional noise for larger studies. Both of these, along with Freelancer, are also recommended for security studies. We conclude the paper by discussing the impact of our results on future security development studies.

The monolithic programming model has been favored for high compatibility and easing the programming for SGX enclaves, i.e., running the secure code with all dependent libraries or even library OSes (LibOSes). Yet, it inevitably bloats the trusted computing base (TCB) and thus deviates from the goal of high security. Introducing fine-grained isolation can effectively mitigate TCB bloating while existing solutions face performance issues. We observe that the off-the-shelf Intel MPK is a perfect match for efficient intra-enclave isolation. Nonetheless, the trust models between MPK and SGX are incompatible by design. We hence propose LIGHTENCLAVE, which embraces non-intrusive extensions on existing SGX hardware to incorporate MPK securely and allows multiple light-enclaves isolated within one enclave. Experiments show that LIGHTENCLAVE incurs up to 4% overhead when separating secret SSL keys for server applications and can significantly improve the performance of Graphene-SGX and Occlum by reducing the communication and runtime overhead, respectively.

We quantitatively investigated the current state of Password Manager (PM) usage and general password habits at a large, private university in the United States. Building on prior qualitative findings from SOUPS 2019, we survey n=277 faculty, staff, and students, finding that 77% of our participants already use PMs, but users of third-party PMs, as opposed to browser-based PMs, were significantly less likely to reuse their passwords across accounts. The largest factor encouraging PM adoption is perceived ease-of-use, indicating that communication and institutional campaigns should focus more on usability factors. Additionally, our work indicates the need for design improvements for browser-based PMs to encourage less password reuse as they are more widely adopted.

Universal cross-site scripting (UXSS) is a browser vulnerability, making a vulnerable browser execute an attacker's script on any web pages loaded by the browser. UXSS is considered a far more severe vulnerability than well-studied cross-site scripting (XSS). This is because the impact of UXSS is not limited to a web application, but it impacts each and every web application as long as a victim user runs a vulnerable browser. We find that UXSS vulnerabilities are difficult to find, especially through fuzzing, for the following two reasons. First, it is challenging to detect UXSS because it is a semantic vulnerability. In order to detect UXSS, one needs to understand the complex interaction semantics between web pages. Second, it is difficult to generate HTML inputs that trigger UXSS since one needs to drive the browser to perform complex interactions and navigations.

This paper proposes FuzzOrigin, a browser fuzzer designed to detect UXSS vulnerabilities. FuzzOrigin addresses the above two challenges by (i) designing an origin sanitizer with a static origin tagging mechanism and (ii) prioritizing origin-update operations through generating chained-navigation operations handling dedicated events. We implemented FuzzOrigin, which works with most modern browsers, including Chrome, Firefox, Edge, and Safari. During the evaluation, FuzzOrigin discovered four previously unknown UXSS vulnerabilities, one in Chrome and three in Firefox, all of which have been confirmed by the vendors. FuzzOrigin is responsible for finding one out of two UXSS vulnerabilities in Chrome reported in 2021 and all three in Firefox, highlighting its strong effectiveness in finding new UXSS vulnerabilities.

ARM is becoming more popular in desktops and data centers, opening a new realm in terms of security attacks against ARM. ARM has released Pointer Authentication, a new hardware security feature that is intended to ensure pointer integrity with cryptographic primitives.

In this paper, we utilize Pointer Authentication (PA) to build a novel scheme to completely prevent any misuse of security-sensitive pointers. We propose PACTIGHT to tightly seal these pointers. PACTIGHT utilizes a strong and unique modifier that addresses the current issues with the state-of-the-art PA defense mechanisms. We implement four defenses based on the PACTIGHT mechanism. Our security and performance evaluation results show that PACTIGHT defenses are more efficient and secure. Using real PA instructions, we evaluated PACTIGHT on 30 different applications, including NGINX web server, with an average performance overhead of 4.07% even when enforcing our strongest defense. PACTIGHT demonstrates its effectiveness and efficiency with real PA instructions on real hardware.

Section 702 of the Foreign Intelligence Surveillance Act authorizes U.S. intelligence agencies to intercept communications content without obtaining a warrant. While Section 702 requires targeting foreigners abroad for intelligence purposes, agencies "incidentally" collect communications to or from Americans and can search that data for purposes beyond intelligence gathering. For over a decade, members of Congress and civil society organizations have called on the U.S. Intelligence Community (IC) to estimate the scale of incidental collection. Senior intelligence officials have acknowledged the value of quantitative transparency for incidental collection, but the IC has not identified a satisfactory estimation method that respects individual privacy, protects intelligence sources and methods, and imposes minimal burden on IC resources.

In this work, we propose a novel approach to estimating incidental collection using secure multiparty computation (MPC). The IC possesses records about the parties to intercepted communications, and communications services possess country-level location for users. By combining these datasets with MPC, it is possible to generate an automated aggregate estimate of incidental collection that maintains confidentiality for intercepted communications and user locations.

We formalize our proposal as a new variant of private set intersection, which we term multiparty private set intersection with union and sum (MPSIU-Sum). We then design and evaluate an efficient MPSIU-Sum protocol, based on elliptic curve cryptography and partially homomorphic encryption. Our protocol performs well at the large scale necessary for estimating incidental collection in Section 702 surveillance.

Payment channel networks (PCNs) provide a faster and cheaper alternative to transactions recorded on the blockchain. Clients can trustlessly establish payment channels with relays by locking coins and then send signed payments that shift coin balances over the network's channels. Although payments are never published, anyone can track a client's payment by monitoring changes in coin balances over the network's channels. We present Twilight, the first PCN that provides a rigorous differential privacy guarantee to its users. Relays in Twilight run a noisy payment processing mechanism that hides the payments they carry. This mechanism increases the relay's cost, so Twilight combats selfish relays that wish to avoid it, using a trusted execution environment (TEE) that ensures they follow its protocol. The TEE does not store the channel's state, which minimizes the trusted computing base. Crucially, Twilight ensures that even if a relay breaks the TEE's security, it cannot break the integrity of the PCN. We analyze Twilight in terms of privacy and cost and study the trade-off between them. We implement Twilight using Intel's SGX framework and evaluate its performance using relays deployed on two continents. We show that a route consisting of 4 relays handles 820 payments/sec.

Hardware flaws are permanent and potent: hardware cannot be patched once fabricated, and any flaws may undermine even formally verified software executing on top. Consequently, verification time dominates implementation time. The gold standard in hardware Design Verification (DV) is dynamic random testing, due to its scalability to large designs. However, given its undirected nature, this technique is inefficient.

Instead of making incremental improvements to existing dynamic hardware verification approaches, we leverage the observation that existing software fuzzers already provide such a solution, and hence adapt them for hardware verification. Specifically, we translate RTL hardware to a software model and fuzz that model directly. The central challenge we address is how to mitigate the differences between the hardware and software execution models. This includes: 1) how to represent test cases, 2) what is the hardware equivalent of a crash, 3) what is an appropriate coverage metric, and 4) how to create a general-purpose fuzzing harness for hardware.

To evaluate our approach, we design, implement, and open-source a Hardware Fuzzing Pipeline that enables fuzzing hardware at scale, using only open-source tools. Using our pipeline, we fuzz five IP blocks from Google's OpenTitan Root-of-Trust chip, four SiFive TileLink peripherals, three RISC-V CPUs, and an FFT accelerator. Our experiments reveal a two orders-of-magnitude reduction in run time to achieve similar Finite State Machine coverage over traditional dynamic verification schemes, and 26.70% better HDL line coverage than prior work. Moreover, with our bus-centric harness, we achieve over 83% HDL line coverage in four of the five OpenTitan IPs we study—without any initial seeds—and are able to detect all bugs (four synthetic from Hack@DAC and one real) implanted across all five OpenTitan IPs we study, with less than 10 hours of fuzzing.

Secure two-party protocols that compute intersection-related statistics have attracted much attention from the industry. These protocols enable two organizations to jointly compute a function (e.g., count and sum) over the intersection of their sets without explicitly revealing this intersection. However, most of such protocols will reveal the intersection size of the two sets in the end. In this work, we are interested in how well an attacker can leverage the revealed intersection sizes to infer some elements' membership of one organization's set. Even disclosing an element's membership of one organization's set to the other organization may violate privacy regulations (e.g., GDPR) since such an element is usually used to identify a person between two organizations. We are the first to study this set membership leakage in intersection-size-revealing protocols. We propose two attacks, namely, baseline attack and feature-aware attack, to evaluate this leakage in realistic scenarios. In particular, our feature-aware attack exploits the realistic set bias that elements with specific features are more likely to be the members of one organization's set. The results show that our two attacks can infer 2.0 ∼ 72.7 set members on average in three realistic scenarios. If the set bias is not weak, the feature-aware attack will outperform the baseline one. For example, in COVID-19 contact tracing, the feature-aware attack can find 25.9 tokens of infected patients in 135 protocol invocations, 1.5 × more than the baseline attack. We discuss how such results may cause negative real-world impacts and propose possible defenses against our attacks.

Smart home devices transmit highly sensitive usage information to servers owned by vendors or third-parties as part of their core functionality. Hence, it is necessary to provide users with the context in which their device data is collected and shared, to enable them to weigh the benefits of deploying smart home technology against the resulting loss of privacy. As privacy policies are generally expected to precisely convey this information, we perform a systematic and data-driven analysis of the current state of smart home privacy policies, with a particular focus on three key questions: (1) how hard privacy policies are for consumers to obtain, (2) how existing policies describe the collection and sharing of device data, and (3) how accurate these descriptions are when compared to information derived from alternate sources. Our analysis of 596 smart home vendors, affecting 2, 442 smart home devices yields 17 findings that impact millions of users, demonstrate gaps in existing smart home privacy policies, as well as challenges and opportunities for automated analysis.

The vulnerability of deep neural networks (DNN) to backdoor (trojan) attacks is extensively studied for the image domain. In a backdoor attack, a DNN is modified to exhibit expected behaviors under attacker-specified inputs (i.e., triggers). Exploring the backdoor vulnerability of DNN in natural language processing (NLP), recent studies are limited to using specially added words/phrases as the trigger pattern (i.e., word-based triggers), which distorts the semantics of the base sentence, causes perceivable abnormality in linguistic features and can be eliminated by potential defensive techniques.

In this paper, we present LiMnguistic Style-Motivated backdoor attack (LISM), the first hidden trigger backdoor attack which exploits implicit linguistic styles for backdooring NLP models. Besides the basic requirements on attack success rate and normal model performance, LISM realizes the following advanced design goals compared with previous word-based backdoor: (a) LISM weaponizes text style transfer models to learn to generate sentences with an attacker-specified linguistic style (i.e., trigger style), which largely preserves the malicious semantics of the base sentence and reveals almost no abnormality exploitable by detection algorithms. (b) Each base sentence is dynamically paraphrased to hold the trigger style, which has almost no dependence on common words or phrases and therefore evades existing defenses which exploit the strong correlation between trigger words and misclassification. Extensive evaluation on 5 popular model architectures, 3 real-world security-critical tasks, 3 trigger styles and 3 potential countermeasures strongly validates the effectiveness and the stealthiness of LISM.

Since its creation, Certificate Transparency (CT) has served as a vital component of the secure web. However, with the increase in TLS adoption, CT has essentially become a defacto log for all newly-created websites, announcing to the public the existence of web endpoints, including those that could have otherwise remained hidden. As a result, web bots can use CT to probe websites in real time, as they are created. Little is known about these bots, their behaviors, and their intentions.

In this paper we present CTPOT, a distributed honeypot system which creates new TLS certificates for the purpose of advertising previously non-existent domains, and records the activity generated towards them from a number of network vantage points. Using CTPOT, we create 4,657 TLS certificates over a period of ten weeks, attracting 1.5 million web requests from 31,898 unique IP addresses. We find that CT bots occupy a distinct subset of the overall web bot population, with less than 2% overlap between IP addresses of CT bots and traditional host-scanning web bots. By creating certificates with varying content types, we are able to further sub-divide the CT bot population into subsets of varying intentions, revealing a stark contrast in malicious behavior among these groups. Finally, we correlate observed bot IP addresses into campaigns using the file paths requested by each bot, and find 105 malicious campaigns targeting the domains we advertise. Our findings shed light onto the CT bot ecosystem, revealing that it is not only distinct to that of traditional IP-based bots, but is composed of numerous entities with varying targets and behaviors.

A number of recent studies have investigated online anonymous ("dark web") marketplaces. Almost all leverage a "measurement-by-proxy" design, in which researchers scrape market public pages, and take buyer reviews as a proxy for actual transactions, to gain insights into market size and revenue. Yet, we do not know if and how this method biases results.

We build a framework to reason about marketplace measurement accuracy, and use it to contrast estimates projected from scrapes of Hansa Market with data from a back-end database seized by the police. We further investigate, by simulation, the impact of scraping frequency, consistency and rate-limits. We find that, even with a decent scraping regimen, one might miss approximately 46% of objects—with scraped listings differing significantly from not-scraped listings on price, views and product categories. This bias also impacts revenue calculations. We find Hansa's total market revenue to be US $50M, which projections based on our scrapes underestimate by a factor of four. Simulations further show that studies based on one or two scrapes are likely to suffer from a very poor coverage (on average, 14% to 30%, respectively).

A high scraping frequency is crucial to achieve reliable coverage, even without a consistent scraping routine. When high-frequency scraping is difficult, e.g., due to deployed anti-scraping countermeasures, innovative scraper design, such as scraping most popular listings first, helps improve coverage. Finally, abundance estimators can provide insights on population coverage when population sizes are unknown.

CPU vulnerabilities undermine the security guarantees provided by software- and hardware-security improvements. While the discovery of transient-execution attacks increased the interest in CPU vulnerabilities on a microarchitectural level, architectural CPU vulnerabilities are still understudied.

In this paper, we systematically analyze existing CPU vulnerabilities showing that CPUs suffer from vulnerabilities whose root causes match with those in complex software. We show that transient-execution attacks and architectural vulnerabilities often arise from the same type of bug and identify the blank spots. Investigating the blank spots, we focus on architecturally improperly initialized data locations.

We discover ÆPIC Leak, the first architectural CPU bug that leaks stale data from the microarchitecture without using a side channel. ÆPIC Leak works on all recent Sunny- Cove-based Intel CPUs (i.e., Ice Lake and Alder Lake). It architecturally leaks stale data incorrectly returned by reading undefined APIC-register ranges. ÆPIC Leak samples data transferred between the L2 and last-level cache, including SGX enclave data, from the superqueue. We target data in use, e.g., register values and memory loads, as well as data at rest, e.g., SGX-enclave data pages. Our end-to-end attack extracts AES-NI, RSA, and even the Intel SGX attestation keys from enclaves within a few seconds. We discuss mitigations and conclude that the only short-term mitigations for ÆPIC Leak are to disable APIC MMIO or not rely on SGX.

Secure multi-party computation (MPC) is an essential tool for privacy-preserving machine learning (ML). However, secure training of large-scale ML models currently requires a prohibitively long time to complete. Given that large ML inference and training tasks in the plaintext setting are significantly accelerated by Graphical Processing Units (GPUs), this raises the natural question: can secure MPC leverage GPU acceleration? A few recent works have studied this question in the context of accelerating specific components or protocols, but do not provide a general-purpose solution. Consequently, MPC developers must be both experts in cryptographic protocol design and proficient at low-level GPU kernel development to achieve good performance on any new protocol implementation.

We present Piranha, a general-purpose, modular platform for accelerating secret sharing-based MPC protocols using GPUs. Piranha allows the MPC community to easily leverage the benefits of a GPU without requiring GPU expertise. Piranha contributes a three-layer architecture: (1) a device layer that can independently accelerate secret-sharing protocols by providing integer-based kernels absent in current general-purpose GPU libraries, (2) a modular protocol layer that allows developers to maximize utility of limited GPU memory with in-place computation and iterator-based support for non-standard memory access patterns, and (3) an application layer that allows applications to remain completely agnostic to the underlying protocols they use.

To demonstrate the benefits of Piranha, we implement 3 state-of-the-art linear secret sharing MPC protocols for secure NN training: 2-party SecureML (IEEE S&P '17), 3-party Falcon (PETS '21), and 4-party FantasticFour (USENIX Security '21). Compared to their CPU-based implementations, the same protocols implemented on top of Piranha's protocol-agnostic acceleration exhibit a 16-48x decrease in training time. For the first time, Piranha demonstrates the feasibility of training a realistic neural network (e.g. VGG), end-to-end, using MPC in a little over one day. Piranha is open source and available at https://github.com/ucbrise/piranha.

Microarchitectural side channels are a pressing security threat. These channels are created when programs modulate hardware resources in a secret data-dependent fashion. They are broadly classified as being either stateful or stateless (also known as contention-based), depending on whether they leave behind a trace for attackers to later observe. Common wisdom suggests that stateful channels are significantly easier to monitor than stateless ones, and hence have received the most attention.

In this paper, we present a novel stateless attack that shows this common wisdom is not always true. Our attack, called Binoculars, exploits unexplored interactions between in-flight page walk operations and other memory operations. Unlike other stateless channels, Binoculars creates significant timing perturbations—up to 20,000 cycles stemming from a single dynamic instruction—making it easy to monitor. We show how these perturbations are address dependent, enabling Binoculars to leak more virtual address bits in victim memory operations than any prior channel. Binoculars needs no shared memory between the attacker and the victim.

Using Binoculars, we design both covert- and side-channel attacks. Our covert channel achieves a high capacity of 1116 KB/s on a Cascade Lake-X machine. We then design a sidechannel attack that steals keys from OpenSSL's side-channel resistant ECDSA by learning the ECDSA nonce k. Binoculars' ability to significantly amplify subtle behaviors, e.g., orderings of stores, is crucial for this attack to succeed because the nonce changes after each run. Finally, we fully break kernel ASLR.

Vulnerabilities inherited from third-party open-source software (OSS) components can compromise the entire software security. However, discovering propagated vulnerable code is challenging as it proliferates with various code syntaxes owing to the OSS modifications, more specifically, internal (e.g., OSS updates) and external modifications of OSS (e.g., code changes that occur during the OSS reuse).

In this paper, we present MOVERY, a precise approach for discovering vulnerable code clones (VCCs) from modified OSS components. By considering the oldest vulnerable function and extracting only core vulnerable and patch lines from security patches, MOVERY generates vulnerability and patch signatures that effectively address OSS modifications. For scalability, MOVERY reduces the search space of the target software by focusing only on the codes borrowed from other OSS projects. Finally, MOVERY determines that the function is VCC when it matches the vulnerability signature and is distinctive from the patch signature.

When we applied MOVERY on ten popular software selected from diverse domains, we observed that 91% of the discovered VCCs had different code syntax from the disclosed vulnerable function. Nonetheless, MOVERY discovered VCCs at least 2.5 times more than those discovered in existing techniques, with much higher accuracy: MOVERY discovered 415 VCCs with 96% precision and 96% recall, whereas two recent VCC discovery techniques, which hardly consider internal and external OSS modifications, discovered only 163 and 72 VCCs with at most 77% precision and 38% recall.

Many protocol implementations are reactive systems, where the protocol process is in continuous interaction with other processes and the environment. If a bug can be exposed only in a certain state, a fuzzer needs to provide a specific sequence of events as inputs that would take protocol into this state before the bug is manifested. We call these bugs as "stateful" bugs. Usually, when we are testing a protocol implementation, we do not have a detailed formal specification of the protocol to rely upon. Without knowledge of the protocol, it is inherently difficult for a fuzzer to discover such stateful bugs. A key challenge then is to cover the state space without an explicit specification of the protocol. Finding stateful bugs in protocol implementations would thus involve partially uncovering the state space of the protocol. Fuzzing stateful software systems would need to incorporate strategies for state identification. Such state identification may follow from manual guidance, or from automatic analysis.

In this work, we posit that manual annotations for state identification can be avoided for stateful protocol fuzzing. Specifically, we rely on a programmatic intuition that the state variables used in protocol implementations often appear in enum type variables whose values (the state names) come from named constants. In our analysis of the Top-50 most widely used open-source protocol implementations, we found that every implementation uses state variables that are assigned named constants (with easy to comprehend names such as INIT, READY) to represent the current state. In this work, we propose to automatically identify such state variables and track the sequence of values assigned to them during fuzzing to produce a "map" of the explored state space.

Our experiments confirm that our stateful fuzzer discovers stateful bugs twice as fast as the baseline greybox fuzzer that we extended. Starting from the initial state, our fuzzer exercises one order of magnitude more state/transition sequences and covers code two times faster than the baseline fuzzer. Several zero-day bugs in prominent protocol implementations were found by our fuzzer, and 8 CVEs have been assigned.

Today, there is limited knowledge about the behavior of UAVs under GPS spoofing attacks in a real-world environment, in particular considering the interplay between the UAV's software as well as other equipped navigation aids and vision sensors. This work aims to understand the feasibility and requirements of fully controlling a UAV's movements by spoofing GPS signals alone. We enumerate the challenges in accomplishing a complete UAV takeover through GPS spoofing and controlling it without crashing. We design and implement a Real-time GPS Signal Generator (RtGSG) that can be configured to generate any arbitrary trajectory and is capable of making changes to GPS signals in real-time through user input, e.g., using a keyboard or joystick. We evaluate RtGSG on popular commercial UAVs from DJI and Autel through over-the-air spoofing experiments in a controlled chamber. We explore generic and UAV-specific GPS spoofing strategies in order to best achieve complete maneuvering control (e.g., velocity and direction). This work highlights that, although COTS UAVs remain vulnerable to GPS spoofing attacks, a complete takeover and control of the UAV requires careful manipulation of the spoofing signals in real-time. Finally, we release our implementation to the scientific community for further research.

Intel's Software Guard Extensions (SGX) provide a nonintrospectable trusted execution environment (TEE) to protect security-critical code from a potentially malicious OS. This protection can only be effective if the individual enclaves are secure, which is already challenging in regular software, and this becomes even more difficult for enclaves as the entire environment is potentially malicious. As such, many enclaves expose common vulnerabilities, e.g., memory corruption and SGXspecific vulnerabilities like null-pointer dereferences. While fuzzing is a popular technique to assess the security of software, dynamically analyzing enclaves is challenging as enclaves are meant to be non-introspectable. Further, they expect an allocated multi-pointer structure as input instead of a plain buffer.

In this paper, we present SGXFUZZ, a coverage-guided fuzzer that introduces a novel binary input structure synthesis method to expose enclave vulnerabilities even without source-code access. To obtain code coverage feedback from enclaves, we show how to extract enclave code from distribution formats. We also present an enclave runner that allows execution of the extracted enclave code as a user-space application at native speed, while emulating all relevant environment interactions of the enclave. We use this setup to fuzz enclaves using a state-of-the-art snapshot fuzzing engine that deploys our novel structure synthesis stage. This stage synthesizes multi-layer pointer structures and size fields incrementally on-the-fly based on fault signals. Furthermore, it matches the expected input format of the enclave without any prior knowledge. We evaluate our approach on 30 open- and closed-source enclaves and found a total of 79 new bugs and vulnerabilities.

Contrastive learning pre-trains an image encoder using a large amount of unlabeled data such that the image encoder can be used as a general-purpose feature extractor for various downstream tasks. In this work, we propose PoisonedEncoder, a data poisoning attack to contrastive learning. In particular, an attacker injects carefully crafted poisoning inputs into the unlabeled pre-training data, such that the downstream classifiers built based on the poisoned encoder for multiple target downstream tasks simultaneously classify attacker-chosen, arbitrary clean inputs as attacker-chosen, arbitrary classes. We formulate our data poisoning attack as a bilevel optimization problem, whose solution is the set of poisoning inputs; and we propose a contrastive-learning-tailored method to approximately solve it. Our evaluation on multiple datasets shows that PoisonedEncoder achieves high attack success rates while maintaining the testing accuracy of the downstream classifiers built upon the poisoned encoder for non-attacker-chosen inputs. We also evaluate five defenses against PoisonedEncoder, including one pre-processing, three in-processing, and one post-processing defenses. Our results show that these defenses can decrease the attack success rate of PoisonedEncoder, but they also sacrifice the utility of the encoder or require a large clean pre-training dataset.

System logs are crucial for forensic analysis, but to be useful, they need to be tamper-proof. To protect the logs, a number of secure logging systems have been proposed from both academia and the industry. Unfortunately, except for the recent KennyLoggings construction, all other logging systems are broken by an attack of Paccagnella et al. (CCS 2020). In this work, we build a secure logging system that improves KennyLoggings in several fronts: adoptability, security, and performance. Our key insight for performance gain is to use AES on a fixed, known key. While this trick is widely used in secure distributed computing, this is the first time it has found an application in the area of symmetric-key cryptography.

Business Collaboration Platforms like Microsoft Teams and Slack enable teamwork by supporting text chatting and third-party resource integration. A user can access online file storage, make video calls, and manage a code repository, all from within the platform, thus making them a hub for sensitive communication and resources. The key enabler for these productivity features is a third-party application model. We contribute an experimental security analysis of this model and the third-party apps. Performing this analysis is challenging because commercial platforms and their apps are closed-source systems. Our analysis methodology is to systematically investigate different types of interactions possible between apps and users. We discover that the access control model in these systems violates two fundamental security principles: least privilege and complete mediation. These violations enable a malicious app to exploit the confidentiality and integrity of user messages and third-party resources connected to the platform. We construct proof-of-concept attacks that can: (1) eavesdrop on user messages without having permission to read those messages; (2) launch fake video calls; (3) automatically merge code into repositories without user approval or involvement. Finally, we provide an analysis of countermeasures that systems like Slack and Microsoft Teams can adopt today.

Behavioral data generated by users’ devices, ranging from emoji use to pages visited, are collected at scale to improve apps and services. These data, however, contain fine-grained records and can reveal sensitive information about individual users. Local differential privacy has been used by companies as a solution to collect data from users while preserving privacy. We here first introduce pool inference attacks, where an adversary has access to a user’s obfuscated data, defines pools of objects, and exploits the user’s polarized behavior in multiple data collections to infer the user’s preferred pool. Second, we instantiate this attack against Count Mean Sketch, a local differential privacy mechanism proposed by Apple and deployed in iOS and Mac OS devices, using a Bayesian model. Using Apple’s parameters for the privacy loss ε, we then consider two specific attacks: one in the emojis setting— where an adversary aims at inferring a user’s preferred skin tone for emojis — and one against visited websites — where an adversary wants to learn the political orientation of a user from the news websites they visit. In both cases, we show the attack to be much more effective than a random guess when the adversary collects enough data. We find that users with high polarization and relevant interest are significantly more vulnerable, and we show that our attack is well-calibrated, allowing the adversary to target such vulnerable users. We finally validate our results for the emojis setting using user data from Twitter. Taken together, our results show that pool inference attacks are a concern for data protected by local differential privacy mechanisms with a large ε, emphasizing the need for additional technical safeguards and the need for more research on how to apply local differential privacy for multiple collections.

HTTP/2 adoption is rapidly climbing. However, in practice, Internet communications still rarely happen over end-to-end HTTP/2 channels. This is due to Content Delivery Networks and other reverse proxies, ubiquitous and necessary components of the Internet ecosystem, which only support HTTP/2 on the client's end, but not the forward connection to the origin server. Instead, proxy technologies predominantly rely on HTTP/2-to-HTTP/1 protocol conversion between the two legs of the connection.

We present the first systematic exploration of HTTP/2-to-HTTP/1 protocol conversion anomalies and their security implications. We develop a novel grammar-based fuzzer for HTTP/2, experiment with 12 popular reverse proxy technologies & CDNs through HTTP/2 frame sequence and content manipulation, and discover a plethora of novel web application attack vectors that lead to Request Blackholing, Denial-of-Service, Query-of-Death, and Request Smuggling attacks.

In November 2020, Antrim County, Michigan published unofficial election results that misstated totals in the presidential race and other contests by up to several thousand votes. Antrim subsequently issued a series of corrections, and the certified presidential results were confirmed by a hand count. Nevertheless, Antrim was repeatedly cited by the former President as evidence of widespread fraud, and it remains a centerpiece of conspiracy theories about the 2020 election. At the request of the Michigan Secretary of State and Attorney General, I performed a forensic investigation of the incident. Using data from the election system, I precisely reproduce the major anomalies, explain their cause, and verify they have been corrected. However, I also uncover other errors affecting specific down-ballot contests that have not been corrected, despite the unusual attention focused on the results, one of which may have changed the outcome of a local contest. Based on this analysis, I refute false claims and disinformation about the incident, concluding that it was not the result of a security breach but rather a series of operator errors compounded by inadequate procedures and insufficiently defensive software design. These events offer lessons for election administration and highlight the value of rigorously investigating election technology incidents for enhancing accuracy and public trust.

Privacy and security challenges due to the outsourcing of data storage and processing to third-party cloud providers are well known. With regard to data privacy, Oblivious RAM (ORAM) schemes provide strong privacy guarantees by not only hiding the contents of the data (by encryption) but also obfuscating the access patterns of the outsourced data. But most existing ORAM datastores are not fault tolerant in that if the external storage server (which stores encrypted data) or the trusted proxy (which stores the encryption key and other metadata) crashes, an application loses all of its data. To achieve fault tolerance, we propose QuORAM, the first ORAM datastore to replicate data with a quorum-based replication protocol. QuORAM's contributions are three-fold: (i) it obfuscates access patterns to provide obliviousness guarantees, (ii) it replicates data using a novel lock-free and decentralized replication protocol to achieve fault tolerance, and (iii) it guarantees linearizable semantics. Experimentally evaluating QuORAM highlights counter-intuitive results: QuORAM incurs negligible cost to achieve obliviousness when compared to an insecure fault-tolerant replicated system; QuORAM's peak throughput is 2.4x of its non-replicated baseline; and QuORAM performs 33.2x better in terms of throughput than an ORAM datastore that relies on CockroachDB, an open-source geo-replicated database, for fault tolerance.

The usage of Deep Neural Networks (DNNs) has steadily increased in recent years. Especially when used in edge devices, dedicated DNN compilers are used to compile DNNs into binaries. Many security applications (such as DNN model extraction, white-box adversarial sample generation, and DNN model patching and hardening) are possible when a DNN model is accessible. However, these techniques cannot be applied to compiled DNNs. Unfortunately, no dedicated decompiler exists that is able to recover a high-level representation of a DNN starting from its compiled binary code.

To address this issue, we propose DnD, the first compiler- and ISA-agnostic DNN decompiler. DnD uses symbolic execution, in conjunction with a dedicated loop analysis, to lift the analyzed binary code into a novel intermediate representation, able to express the high-level mathematical DNN operations in a compiler- and ISA-agnostic way. Then, DnD matches the extracted mathematical DNN operations with template mathematical DNN operations, and it recovers hyper-parameters and parameters of all the identified DNN operators, as well as the overall DNN topology. Our evaluation shows that DnD can perfectly recover different DNN models, extracting them from binaries compiled by two different compilers (Glow and TVM) for three different ISAs (Thumb, AArch64, and x86-64). Moreover, DnD enables extracting the DNN models used by real-world micro-controllers and attacking them using white-box adversarial machine learning techniques.

End-to-end encryption provides strong privacy protections to billions of people, but it also complicates efforts to moderate content that can seriously harm people. To address this concern, Tyagi et al. [CRYPTO 2019] introduced the concept of asymmetric message franking (AMF) so that people can report abusive content to a moderator, while otherwise retaining end-to-end privacy by default and compatibility with anonymous communication systems like Signal's sealed sender.

In this work, we provide a new construction for asymmetric message franking called Hecate that is faster, more secure, and introduces additional functionality compared to Tyagi et al. First, our construction uses fewer invocations of standardized crypto primitives and operates in the plain model. Second, on top of AMF's accountability and deniability requirements, we also add forward and backward secrecy. Third, we combine AMF with source tracing, another approach to content moderation that has previously been considered only in the setting of non-anonymous networks. Source tracing allows for messages to be forwarded, and a report only identifies the original source who created a message. To provide anonymity for senders and forwarders, we introduce a model of AMF with preprocessing whereby every client authenticates with the moderator out-of-band to receive a token that they later consume when sending a message anonymously.

Dynamic Information Flow Tracking (DIFT) forms the foundation of a wide range of security and privacy analyses. The main challenges faced by DIFT techniques are performance and scalability. Due to the large number of states in a program, the number of data flows can be prohibitively large and efficiently performing interactive data flow analysis queries using existing approaches is challenging. In this paper, we identify that DIFT under dependency-based information flow rules can be cast as linear transformations over taint state. This enables a novel matrix-based representation, which we call FlowMatrix, to represent DIFT operations concisely and makes it practical to adopt GPUs as co-processors for DIFT analysis. FlowMatrix provides efficient support for interactive DIFT query operations. We design a DIFT query system and prototype it on commodity GPUs. Our evaluation shows that our prototype outperforms CPU-based baseline by 5.6 times and enables rapid response to a DIFT queries. It has two to three orders of magnitude higher throughput compared to typical DIFT analysis solutions. We also demonstrate the efficiency and efficacy of new DIFT query operations.

Database management systems (DBMSs) are critical components of modern data-intensive applications. Developers have adopted many testing techniques to detect DBMS bugs such as crashes and assertion failures. However, most previous efforts cannot detect logical bugs that make the DBMS return incorrect results. Recent work proposed several oracles to identify incorrect results, but they rely on rule-based expression generation to synthesize queries without any guidance.

In this paper, we propose to combine coverage-based guidance, validity-oriented mutations and oracles to detect logical bugs in DBMS systems. Specifically, we first design a set of general APIs to decouple the logic of fuzzers and oracles, so that developers can easily port fuzzing tools to test DBMSs and write new oracles for existing fuzzers. Then, we provide validity-oriented mutations to generate high-quality query statements in order to find more logical bugs. Our prototype, SQLRight, outperforms existing tools that only rely on oracles or code coverage. In total, SQLRight detects 18 logical bugs from two well-tested DBMSs, SQLite and MySQL. All bugs have been confirmed and 14 of them have been fixed.

This paper studies microarchitectural side-channel attacks and mitigations on the on-chip mesh interconnect used in modern, server-class Intel processors. We find that, though difficult to exploit, the mesh interconnect can be abused by an adversary even when known attack vectors inside the cores and caches are closed. We then present novel, non-invasive mitigation mechanisms to interconnect side-channel attacks and offer insights to guide the design of future defenses.

Our analysis starts by thoroughly reverse engineering the mesh interconnect to reveal, for the first time, the precise conditions under which it is susceptible to contention. We show that an attacker can use these conditions to build a cross-core covert channel with a capacity of over 1.5 Mbps. We then demonstrate the feasibility of side-channel attacks that leak keys from vulnerable cryptographic implementations by monitoring mesh interconnect contention. Finally, we present an analytical model to quantify the vulnerability levels of different victim and attacker placements on the chip and use the results to design a software-only mitigation mechanism.

Human analysts must reverse engineer binary programs as a prerequisite for a number of security tasks, such as vulnerability analysis, malware detection, and firmware re-hosting. Existing studies of human reversers and the processes they follow are limited in size and often use qualitative metrics that require subjective evaluation.

In this paper, we reframe the problem of reverse engineering binaries as the problem of perfect decompilation, which is the process of recovering, from a binary program, source code that, when compiled, produces binary code that is identical to the original binary. This gives us a quantitative measure of understanding, and lets us examine the reversing process programmatically.

We developed a tool, called Decomperson, that supported a group of reverse engineers during a large-scale security competition designed to collect information about the participants' reverse engineering process, with the well-defined goal of achieving perfect decompilation. Over 150 people participated, and we collected more than 35,000 code submissions, the largest manual reverse engineering dataset to date. This includes snapshots of over 300 successful perfect decompilation attempts. In this paper, we show how perfect decompilation allows programmatic analysis of such large datasets, providing new insights into the reverse engineering process.

It is well known in the cryptographic literature that the most common digital signature schemes used in practice can fail catastrophically in the presence of faults during computation. We use passive and active network measurements to analyze organically-occuring faults in billions of digital signatures generated by tens of millions of hosts.We find that a persistent rate of apparent hardware faults in unprotected implementations has resulted in compromised certificate RSA private keys for years. The faulty signatures we observed allowed us to compute private RSA keys associated with a top-10 Alexa site, several browser-trusted wildcard certificates for organizations that used a popular VPN product, and a small sporadic population of other web sites and network devices. These measurements illustrate the fragility of RSA PKCS#1v1.5 signature padding and provide insight on the risks faced by unprotected implementations on hardware at Internet scale.

Website fingerprinting attacks, which analyse the metadata of encrypted network communication to identify visited websites, have been shown to be effective on privacy-enhancing technologies including virtual private networks (VPNs) and encrypted proxies. Despite this, VPNs are still undefended against these attacks, leaving millions of users vulnerable. Proposed defences against website fingerprinting require cooperation between the client and a remote endpoint to reshape the network traffic, thereby hindering deployment.

We observe that the rapid and wide-spread deployment of QUIC and HTTP/3 creates an exciting opportunity to build website-fingerprinting defences directly into client applications, such as browsers, without requiring any changes to web servers, VPNs, or the deployment of new network services. We therefore design and implement the QCSD framework, which leverages QUIC and HTTP/3 to emulate existing website-fingerprinting defences by bidirectionally adding cover traffic and reshaping connections solely from the client. As case studies, we emulate both the FRONT and Tamaraw defences solely from the client and collected several datasets of live-defended traffic on which we evaluated modern machine-learning based attacks. Our results demonstrate the promise of this approach in shaping connections towards client-orchestrated defences, thereby removing a primary barrier to the deployment of website-fingerprinting defences.

The censorship arms race has recently gone through a transformation, thanks to recent efforts showing that new ways to evade censorship can be discovered in an automated fashion. However, all of these prior automated efforts operate by manipulating TCP/IP headers; while impressive, deploying these have proven challenging, as header modifications often require greater privileges than are available to censorship circumvention apps. In that line of work, the application layer has gone largely unexplored. This is not without reason: the space of application messages is much larger and far less structured than TCP/IP headers.

In this paper, we present the first techniques to automate the discovery of new censorship evasion techniques purely in the application layer. We present a general solution and apply it specifically to HTTP and DNS censorship in China, India, and Kazakhstan. Our automated techniques discovered a total of 77 unique evasion strategies for HTTP and 9 for DNS, all of which require only application-layer modifications, making them easier to incorporate into apps and deploy. We analyze these strategies and shed new light into the inner workings of the censors. We find that the success of application-layer strategies can depend heavily on the type and version of the destination server. Surprisingly, a large class of our evasion strategies exploit instances in which censors are more RFCcompliant than popular application servers. We have made our code publicly available.

Multi-Server Private Information Retrieval (PIR) is a cryptographic protocol that allows a client to securely query a database entry from n ≥ 2 servers of which less than t can collude, s.t. the servers learn no information about the query. Highly efficient PIR could be used for large-scale applications like Compromised Credential Checking (C3) (USENIX Security'19), which allows users to check whether their credentials have been leaked in a data breach. However, state-of-the art PIR schemes are not efficient enough for fast online responses at this scale.

In this work, we introduce Client-Independent Preprocessing (CIP) PIR that moves (t −1)/n of the online computation to a local, client independent, preprocessing phase suitable for efficient batch precomputations. The online performance of CIP-PIR improves linearly with the number of servers n. We show that large-scale applications like C3 with PIR are practical by implementing our CIP-PIR scheme using a parallelized CPU implementation. To the best of our knowledge, this is the first multi-server PIR scheme whose preprocessing phase is completely independent of the client, and where online performance simultaneously improves with the number of servers n. In addition, we accelerate for the first time the huge amount of XOR operations in multi-server PIR with GPUs. Our GPUbased CIP-PIR achieves an improvement up to factor 2.1× over our CPU-based implementation for n = 2 servers, and enables a client to query an entry in a 25 GB database within less than 1 second.