USENIX Security '22 Technical Sessions

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.

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.

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.

We describe a method to separate abuse from legitimate traffic when we have categorical features and no labels are available. Our approach hinges on the observation that, if we could locate them, unattacked bins of a categorical feature x would allow us to estimate the benign distribution of any feature that is independent of x. We give an algorithm that finds these unattacked bins (if they exist) and show how to build an overall classifier that is suitable for very large data volumes and high levels of abuse. The approach is one-sided: our only significant assumptions about abuse are the existence of unattacked bins, and that distributions of abuse traffic do not precisely match those of benign.

We evaluate on two datasets: 3 million requests from a web-server dataset and a collection of 5.1 million Twitter accounts crawled using the public API. The results confirm that the approach is successful at identifying clusters of automated behaviors. On both problems we easily outperform unsupervised methods such as Isolation Forests, and have comparable performance to Botometer on the Twitter dataset.

As users navigate the web they face a multitude of threats; among them, attacks that result in account compromise can be particularly devastating. In a world fraught with data breaches and sophisticated phishing attacks, web services strive to fortify user accounts by adopting new mechanisms that identify and prevent suspicious login attempts. More recently, browser fingerprinting techniques have been incorporated into the authentication workflow of major services as part of their decision-making process for triggering additional security mechanisms (e.g., two-factor authentication).

In this paper we present the first comprehensive and in-depth exploration of the security implications of real-world systems relying on browser fingerprints for authentication. Guided by our investigation, we develop a tool for automatically constructing fingerprinting vectors that replicate the process of target websites, enabling the extraction of fingerprints from users' devices that exactly match those generated by target websites. Subsequently, we demonstrate how phishing attackers can replicate users' fingerprints on different devices to deceive the risk-based authentication systems of high-value web services (e.g., cryptocurrency trading) to completely bypass two-factor authentication. To gain a better understanding of whether attackers can carry out such attacks, we study the evolution of browser fingerprinting practices in phishing websites over time. While attackers do not generally collect all the necessary fingerprinting attributes, unfortunately that is not the case for attackers targeting certain financial institutions where we observe an increasing number of phishing sites capable of pulling off our attacks. To address the significant threat posed by our attack, we have disclosed our findings to the vulnerable vendors.

Automated Teller Machines (ATMs) represent the most used system for withdrawing cash. The European Central Bank reported more than 11 billion cash withdrawals and loading/unloading transactions on the European ATMs in 2019. Although ATMs have undergone various technological evolutions, Personal Identification Numbers (PINs) are still the most common authentication method for these devices. Unfortunately, the PIN mechanism is vulnerable to shoulder-surfing attacks performed via hidden cameras installed near the ATM to catch the PIN pad. To overcome this problem, people get used to covering the typing hand with the other hand. While such users probably believe this behavior is safe enough to protect against mentioned attacks, there is no clear assessment of this countermeasure in the scientific literature.

This paper proposes a novel attack to reconstruct PINs entered by victims covering the typing hand with the other hand. We consider the setting where the attacker can access an ATM PIN pad of the same brand/model as the target one. Afterward, the attacker uses that model to infer the digits pressed by the victim while entering the PIN. Our attack owes its success to a carefully selected deep learning architecture that can infer the PIN from the typing hand position and movements. We run a detailed experimental analysis including 58 users. With our approach, we can guess 30% of the 5-digit PINs within three attempts – the ones usually allowed by ATM before blocking the card. We also conducted a survey with 78 users that managed to reach an accuracy of only 7.92% on average for the same setting. Finally, we evaluate a shielding countermeasure that proved to be rather inefficient unless the whole keypad is shielded.

The ubiquity of user accounts in websites and online services makes account hijacking a serious security concern. Although previous research has studied various techniques through which an attacker can gain access to a victim's account, relatively little attention has been directed towards the process of account creation. The current trend towards federated authentication (e.g., Single Sign-On) adds an additional layer of complexity because many services now support both the classic approach in which the user directly sets a password, and the federated approach in which the user authenticates via an identity provider.

Inspired by previous work on preemptive account hijacking [Ghasemisharif et al., USENIX SEC 2018], we show that there exists a whole class of account pre-hijacking attacks. The distinctive feature of these attacks is that the attacker performs some action before the victim creates an account, which makes it trivial for the attacker to gain access after the victim has created/recovered the account. Assuming a realistic attacker who knows only the victim's email address, we identify and discuss five different types of account pre-hijacking attacks.

To ascertain the prevalence of such vulnerabilities in the wild, we analyzed 75 popular services and found that at least 35 of these were vulnerable to one or more account pre-hijacking attacks. Whilst some of these may be noticed by attentive users, others were completely undetectable from the victim's perspective. Finally, we investigated the root cause of these vulnerabilities and present a set of security requirements to prevent such vulnerabilities arising in future.

Credential stuffing attacks use stolen passwords to log into victim accounts. To defend against these attacks, recently deployed compromised credential checking (C3) services provide APIs that help users and companies check whether a username, password pair is exposed. These services however only check if the exact password is leaked, and therefore do not mitigate credential tweaking attacks — attempts to compromise a user account with variants of a user's leaked passwords. Recent work has shown credential tweaking attacks can compromise accounts quite effectively even when the credential stuffing countermeasures are in place.

We initiate work on C3 services that protect users from credential tweaking attacks. The core underlying challenge is how to identify passwords that are similar to their leaked passwords while preserving honest clients' privacy and also preventing malicious clients from extracting breach data from the service. We formalize the problem and explore ways to measure password similarity that balance efficacy, performance, and security. Based on this study, we design "Might I Get Pwned" (MIGP), a new kind of breach alerting service. Our simulations show that MIGP reduces the efficacy of state-of-the-art 1000-guess credential tweaking attacks by 94%. MIGP preserves user privacy and limits potential exposure of sensitive breach entries. We show that the protocol is fast, with response time close to existing C3 services. We worked with Cloudflare to deploy MIGP in practice.

Passwords remain the primary way to authenticate users online. Yet little is known about the characteristics of login requests submitted to login systems due to the sensitivity of monitoring submitted passwords. This means we don't have answers to basic questions, such as how often users submit a password similar to their actual password, whether users often resubmit the same incorrect password, how many users utilize passwords known to be in a public breach, and more. Whether we can build and deploy measurement infrastructure to safely answer such questions is, itself, an open question.

We offer a system, called Gossamer, that enables securely logging information about login attempts, including carefully chosen statistics about submitted passwords. We provide a simulation-based approach for tuning the security-utility trade-offs for storing different password-derived statistics. This enables us to gather useful measurements while reducing risk even in the unlikely case of complete compromise of the measurement system. We worked closely with two large universities and deployed Gossamer to perform a measurement study that observed 34 million login requests over a seven month period. The measurements we gather provide insight into the use of breached credentials, password usability, and other characteristics of the submitted login requests.

Many applications, from the Web to smart contracts, need to safely execute untrusted code. We observe that WebAssembly (Wasm) is ideally positioned to support such applications, since it promises safety and performance, while serving as a compiler target for many high-level languages. However, Wasm's safety guarantees are only as strong as the implementation that enforces them. Hence, we explore two distinct approaches to producing provably sandboxed Wasm code. One draws on traditional formal methods to produce mathematical, machine-checked proofs of safety. The second carefully embeds Wasm semantics in safe Rust code such that the Rust compiler can emit safe executable code with good performance. Our implementation and evaluation of these two techniques indicate that leveraging Wasm gives us provably-safe multilingual sandboxing with performance comparable to standard, unsafe approaches.

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.

To mitigate a myriad of Web attacks, modern browsers support client-side security policies shipped through HTTP response headers. To enforce these defenses, the server needs to communicate them to the client, a seemingly straightforward process. However, users may access the same site in variegate ways, e.g., using different User-Agents, network access methods, or language settings. All these usage scenarios should enforce the same security policies, otherwise a security lottery would take place: depending on specific client characteristics, different levels of Web application security would be provided to users (inconsistencies). We formalize security guarantees provided through four popular mechanisms and apply this to measure the prevalence of inconsistencies in the security policies of top sites across different client characteristics. Based on our insights, we investigate the security implications of both deterministic and non-deterministic inconsistencies, and show how even prominent services are affected by them.

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.

Web tracking has evolved to become a norm on the Internet. As a matter of fact, the web tracking market has grown to raise billions of dollars. Privacy cautious web practitioners and researchers extensively studied the phenomenon proving how widespread this practice is, and providing effective solutions to give users the option of feeling private while freely surfing the web. However, because all those studies looked at this trend only from the trackers' perspective, still there are a lot of unknowns regarding what the real impact of tracking is on real users. Our goal with this paper is to fill this gap in the web tracking topic. Thanks to logs of web browsing telemetry, we were able to look at this trend from the users' eyes. Precisely, we measure how fast a user encounters trackers and research on options to reduce her privacy risk. Moreover, we also estimate the fraction of browsing histories that are known by trackers and discuss two tracking strategies to increase the existing knowledge about users.

Sender-anonymous end-to-end encrypted messaging allows sending messages to a recipient without revealing the sender's identity to the messaging platform. Signal recently introduced a sender anonymity feature that includes an abuse mitigation mechanism meant to allow the platform to block malicious senders on behalf of a recipient.

We explore the tension between sender anonymity and abuse mitigation. We start by showing limitations of Signal's deployed mechanism, observing that it results in relatively weak anonymity properties and showing a new griefing attack that allows a malicious sender to drain a victim's battery. We therefore design a new protocol, called Orca, that allows recipients to register a privacy-preserving blocklist with the platform. Without learning the sender's identity, the platform can check that the sender is not on the blocklist and that the sender can be identified by the recipient. We construct Orca using a new type of group signature scheme, for which we give formal security notions. Our prototype implementation showcases Orca's practicality.

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.

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.

Modern disassembly tools often rely on empirical evaluations to validate their performance and discover their limitations, thus promoting long-term evolvement. To support the empirical evaluation, a foundation is the right approach to collect the ground truth knowledge. However, there has been no unanimous agreement on the approach we should use. Most users pick an approach based on their experience or will, regardless of the properties that the approach presents.

In this paper, we perform a study on the approaches to building the ground truth for binary disassembly, aiming to shed light on the right way for the future. We first provide a taxonomy of the approaches used by past research, which unveils five major mechanisms behind those approaches. Following the taxonomy, we summarize the properties of the five mechanisms from two perspectives: (i) the coverage and precision of the ground truth produced by the mechanisms and (ii) the applicable scope of the mechanisms (e.g., what disassembly tasks and what types of binaries are supported). The summarization, accompanied by quantitative evaluations, illustrates that many mechanisms are ill-suited to support the generation of disassembly ground truth. The mechanism best serving today's need is to trace the compiling process of the target binaries to collect the ground truth information.

Observing that the existing tool to trace the compiling process can still miss ground truth results and can only handle x86/x64 binaries, we extend the tool to avoid overlooking those results and support ARM32/AArch64/MIPS32/MIPS64 binaries. We envision that our extension will make the tool a better foundation to enable universal, standard ground truth for binary disassembly.

Remote direct memory access (RDMA) has gained popularity in cloud datacenters. In RDMA, clients bypass server CPUs and directly read/write remote memory. Recent findings have highlighted a host of vulnerabilities with RDMA, which give rise to attacks such as packet injection, denial of service, and side channel leakage, but RDMA defenses are still lagging behind. As the RDMA datapath bypasses CPU-based software processing, traditional defenses cannot be easily inserted without incurring performance penalty. Bedrock develops a security foundation for RDMA inside the network, leveraging programmable data planes in modern network hardware. It designs a range of defense primitives, including source authentication, access control, as well as monitoring and logging, to address RDMA-based attacks. Bedrock does not incur software overhead to the critical datapath, and delivers native RDMA performance in data transfers. Moreover, Bedrock operates transparently to legacy RDMA systems, without requiring RNIC, OS, or RDMA library changes. We present a comprehensive set of experiments on Bedrock and demonstrate its effectiveness.

ICMP redirect is a mechanism that allows an end host to dynamically update its routing decisions for particular destinations. Previous studies show that ICMP redirect may be exploited by attackers to manipulate the routing of victim traffic. However, it is widely believed that ICMP redirect attacks are not a real-world threat since they can only occur under specific network topologies (e.g., LAN). In this paper, we conduct a systematic study on the legitimacy check mechanism of ICMP and uncover a fundamental gap between the check mechanism and stateless protocols, resulting in a wide range of vulnerabilities. In particular, we find that off-path attackers can utilize a suite of stateless protocols (e.g., UDP, ICMP, GRE, IPIP and SIT) to easily craft evasive ICMP error messages, thus revitalizing ICMP redirect attacks to cause serious damage in the real world, particularly, on the wide-area network. First, we show that off-path attackers can conduct a stealthy DoS attack by tricking various public servers on the Internet into mis-redirecting their traffic into black holes with a single forged ICMP redirect message. For example, we reveal that more than 43K popular websites on the Internet are vulnerable to this DoS attack. In addition, we identify 54.47K open DNS resolvers and 186 Tor nodes on the Internet are vulnerable as well. Second, we show that, by leveraging ICMP redirect attacks against NATed networks, off-path attackers in the same NATed network can perform a man-in-the-middle (MITM) attack to intercept the victim traffic. Finally, we develop countermeasures to throttle the attacks.

When a human activity requires a lot of expertise and very specialized cognitive skills that are poorly understood by the general population, it is often considered `an art.' Different activities in the security domain have fallen in this category, such as exploitation, hacking, and the main focus of this paper: binary reverse engineering (RE).

However, while experts in many areas (ranging from chess players to computer programmers) have been studied by scientists to understand their mental models and capture what is special about their behavior, the `art' of understanding binary code and solving reverse engineering puzzles remains to date a black box.

In this paper, we present a measurement of the different strategies adopted by expert and beginner reverse engineers while approaching the analysis of x86 (dis)assembly code, a typical static RE task. We do that by performing an exploratory analysis of data collected over 16,325 minutes of RE activity of two unknown binaries from 72 participants with different experience levels: 39 novices and 33 experts.

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.

Users rely on ad and tracker blocking tools to protect their privacy. Unfortunately, existing ad and tracker blocking tools are susceptible to mutable advertising and tracking content. In this paper, we first demonstrate that a state-of-the-art ad and tracker blocker, AdGraph, is susceptible to such adversarial evasion techniques that are currently deployed on the web. Second, we introduce WebGraph, the first ML-based ad and tracker blocker that detects ads and trackers based on their action rather than their content. By featurizing the actions that are fundamental to advertising and tracking information flows – e.g., storing an identifier in the browser or sharing an identifier with another tracker – WebGraph performs nearly as well as prior approaches, but is significantly more robust to adversarial evasions. In particular, we show that WebGraph achieves comparable accuracy to AdGraph, while significantly decreasing the success rate of an adversary from near-perfect for AdGraph to around 8% for WebGraph. Finally, we show that WebGraph remains robust to sophisticated adversaries that use adversarial evasion techniques beyond those currently deployed on the web.

Request chains are being used by advertisers and trackers for information sharing and circumventing recently introduced privacy protections in web browsers. There is little prior work on mitigating the increasing exploitation of request chains by advertisers and trackers. The state-of-the-art ad and tracker blocking approaches lack the necessary context to effectively detect advertising and tracking request chains. We propose Khaleesi, a machine learning approach that captures the essential sequential context needed to effectively detect advertising and tracking request chains. We show that Khaleesi achieves high accuracy, that holds well over time, is generally robust against evasion attempts, and outperforms existing approaches. We also show that Khaleesi is suitable for online deployment and it improves page load performance.