Dan Boneh

Andy K. Zhang, Neil Perry, Riya Dulepet et al.

Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause real-world impact. Policymakers, model providers, and researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. We include 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structed bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. All code and data are publicly available at https://cybench.github.io.

Ryan Babbush, Adam Zalcman, Craig Gidney et al.

This whitepaper seeks to elucidate implications that the capabilities of developing quantum architectures have on blockchain vulnerabilities and mitigation strategies. First, we provide new resource estimates for breaking the 256-bit Elliptic Curve Discrete Logarithm Problem, the core of modern blockchain cryptography. We demonstrate that Shor's algorithm for this problem can execute with either <1200 logical qubits and <90 million Toffoli gates or <1450 logical qubits and <70 million Toffoli gates. In the interest of responsible disclosure, we use a zero-knowledge proof to validate these results without disclosing attack vectors. On superconducting architectures with 1e-3 physical error rates and planar connectivity, those circuits can execute in minutes using fewer than half a million physical qubits. We introduce a critical distinction between fast-clock (such as superconducting and photonic) and slow-clock (such as neutral atom and ion trap) architectures. Our analysis reveals that the first fast-clock CRQCs would enable on-spend attacks on public mempool transactions of some cryptocurrencies. We survey major cryptocurrency vulnerabilities through this lens, identifying systemic risks associated with advanced features in some blockchains such as smart contracts, Proof-of-Stake consensus, and Data Availability Sampling, as well as the enduring concern of abandoned assets. We argue that technical solutions would benefit from accompanying public policy and discuss various frameworks of digital salvage to regulate the recovery or destruction of dormant assets while preventing adversarial seizure. We also discuss implications for other digital assets and tokenization as well as challenges and successful examples of the ongoing transition to Post-Quantum Cryptography (PQC). Finally, we urge all vulnerable cryptocurrency communities to join the ongoing migration to PQC without delay.

Erez Badash, Dan Boneh, Ilan Komargodski et al. · microsoft-research, stanford

We present Hawkeye, a system for analyzing and reproducing GPU-level arithmetic operations. Using our framework, anyone can re-execute on a CPU the exact matrix multiplication operations underlying a machine learning model training or inference workflow that was executed on an NVIDIA GPU, without any precision loss. This is in stark contrast to prior approaches to verifiable machine learning, which either introduce significant computation overhead to the original model owner, or suffer from non-robustness and quality degradation. The main technical contribution of Hawkeye is a systematic sequence of carefully crafted tests that study rounding direction, subnormal number handling, and order of (non-associative) accumulation during matrix multiplication on NVIDIA's Tensor Cores. We test and evaluate our framework on multiple NVIDIA GPU architectures ( Ampere, Hopper, and Lovelace) and precision types (FP16, BFP16, FP8). In all test cases, Hawkeye enables perfect reproduction of matrix multiplication on a CPU, paving the way for efficient and trustworthy third-party auditing of ML model training and inference.

Chhavi Yadav, Amrita Roy Chowdhury, Dan Boneh et al.

Machine learning models are increasingly used in societal applications, yet legal and privacy concerns demand that they very often be kept confidential. Consequently, there is a growing distrust about the fairness properties of these models in the minds of consumers, who are often at the receiving end of model predictions. To this end, we propose \name -- a system that uses Zero-Knowledge Proofs (a cryptographic primitive) to publicly verify the fairness of a model, while maintaining confidentiality. We also propose a fairness certification algorithm for fully-connected neural networks which is befitting to ZKPs and is used in this system. We implement \name in Gnark and demonstrate empirically that our system is practically feasible. Code is available at https://github.com/infinite-pursuits/FairProof.

Chhavi Yadav, Evan Monroe Laufer, Dan Boneh et al.

In principle, explanations are intended as a way to increase trust in machine learning models and are often obligated by regulations. However, many circumstances where these are demanded are adversarial in nature, meaning the involved parties have misaligned interests and are incentivized to manipulate explanations for their purpose. As a result, explainability methods fail to be operational in such settings despite the demand \cite{bordt2022post}. In this paper, we take a step towards operationalizing explanations in adversarial scenarios with Zero-Knowledge Proofs (ZKPs), a cryptographic primitive. Specifically we explore ZKP-amenable versions of the popular explainability algorithm LIME and evaluate their performance on Neural Networks and Random Forests. Our code is publicly available at https://github.com/emlaufer/ExpProof.

Justin W. Lin, Eliot Krzysztof Jones, Donovan Julian Jasper et al.

We present the first comprehensive evaluation of AI agents against human cybersecurity professionals in a live enterprise environment. We evaluate ten cybersecurity professionals alongside six existing AI agents and ARTEMIS, our new agent scaffold, on a large university network consisting of ~8,000 hosts across 12 subnets. ARTEMIS is a multi-agent framework featuring dynamic prompt generation, arbitrary sub-agents, and automatic vulnerability triaging. In our comparative study, ARTEMIS placed second overall, discovering 9 valid vulnerabilities with an 82% valid submission rate and outperforming 9 of 10 human participants. While existing scaffolds such as Codex and CyAgent underperformed relative to most human participants, ARTEMIS demonstrated technical sophistication and submission quality comparable to the strongest participants. We observe that AI agents offer advantages in systematic enumeration, parallel exploitation, and cost -- certain ARTEMIS variants cost $18/hour versus $60/hour for professional penetration testers. We also identify key capability gaps: AI agents exhibit higher false-positive rates and struggle with GUI-based tasks.

Megha Srivastava, Simran Arora, Dan Boneh

The increasing compute demands of AI systems have led to the emergence of services that train models on behalf of clients lacking necessary resources. However, ensuring correctness of training and guarding against potential training-time attacks, such as data poisoning and backdoors, poses challenges. Existing works on verifiable training largely fall into two classes: proof-based systems, which are difficult to scale, and ``optimistic'' methods that consider a third-party auditor who can replicate the training process and contest the trainer. A key challenge with the latter is that nondeterminism between GPU types during training prevents exact replication of the training process, resulting in schemes that are non-robust. We propose a method that combines training in a higher precision than the target, rounding after intermediate computations, and sharing rounding decisions based on an adaptive thresholding procedure, to successfully control for nondeterminism. Across three different NVIDIA GPUs (A40, Titan XP, RTX 2080 Ti), we achieve exact training replication at FP32 precision for both full-training and fine-tuning of ResNet-50 (23M) and GPT-2 (117M) models. Our verifiable training scheme significantly decreases the storage and time costs compared to proof-based systems, and is publicly released at https://github.com/meghabyte/verifiable-training.

Florian Tramèr, Pascal Dupré, Gili Rusak et al.

Perceptual ad-blocking is a novel approach that detects online advertisements based on their visual content. Compared to traditional filter lists, the use of perceptual signals is believed to be less prone to an arms race with web publishers and ad networks. We demonstrate that this may not be the case. We describe attacks on multiple perceptual ad-blocking techniques, and unveil a new arms race that likely disfavors ad-blockers. Unexpectedly, perceptual ad-blocking can also introduce new vulnerabilities that let an attacker bypass web security boundaries and mount DDoS attacks. We first analyze the design space of perceptual ad-blockers and present a unified architecture that incorporates prior academic and commercial work. We then explore a variety of attacks on the ad-blocker's detection pipeline, that enable publishers or ad networks to evade or detect ad-blocking, and at times even abuse its high privilege level to bypass web security boundaries. On one hand, we show that perceptual ad-blocking must visually classify rendered web content to escape an arms race centered on obfuscation of page markup. On the other, we present a concrete set of attacks on visual ad-blockers by constructing adversarial examples in a real web page context. For seven ad-detectors, we create perturbed ads, ad-disclosure logos, and native web content that misleads perceptual ad-blocking with 100% success rates. In one of our attacks, we demonstrate how a malicious user can upload adversarial content, such as a perturbed image in a Facebook post, that fools the ad-blocker into removing another users' non-ad content. Moving beyond the Web and visual domain, we also build adversarial examples for AdblockRadio, an open source radio client that uses machine learning to detects ads in raw audio streams.

Saba Eskandarian, Jonathan Cogan, Sawyer Birnbaum et al.

Users regularly enter sensitive data, such as passwords, credit card numbers, or tax information, into the browser window. While modern browsers provide powerful client-side privacy measures to protect this data, none of these defenses prevent a browser compromised by malware from stealing it. In this work, we present Fidelius, a new architecture that uses trusted hardware enclaves integrated into the browser to enable protection of user secrets during web browsing sessions, even if the entire underlying browser and OS are fully controlled by a malicious attacker. Fidelius solves many challenges involved in providing protection for browsers in a fully malicious environment, offering support for integrity and privacy for form data, JavaScript execution, XMLHttpRequests, and protected web storage, while minimizing the TCB. Moreover, interactions between the enclave and the browser, the keyboard, and the display all require new protocols, each with their own security considerations. Finally, Fidelius takes into account UI considerations to ensure a consistent and simple interface for both developers and users. As part of this project, we develop the first open source system that provides a trusted path from input and output peripherals to a hardware enclave with no reliance on additional hypervisor security assumptions. These components may be of independent interest and useful to future projects. We implement and evaluate Fidelius to measure its performance overhead, finding that Fidelius imposes acceptable overhead on page load and user interaction for secured pages and has no impact on pages and page components that do not use its enhanced security features.

David J. Wu, Ankur Taly, Asim Shankar et al.

Automatic service discovery is essential to realizing the full potential of the Internet of Things (IoT). While discovery protocols like Multicast DNS, Apple AirDrop, and Bluetooth Low Energy have gained widespread adoption across both IoT and mobile devices, most of these protocols do not offer any form of privacy control for the service, and often leak sensitive information such as service type, device hostname, device owner's identity, and more in the clear. To address the need for better privacy in both the IoT and the mobile landscape, we develop two protocols for private service discovery and private mutual authentication. Our protocols provide private and authentic service advertisements, zero round-trip (0-RTT) mutual authentication, and are provably secure in the Canetti-Krawczyk key-exchange model. In contrast to alternatives, our protocols are lightweight and require minimal modification to existing key-exchange protocols. We integrate our protocols into an existing open-source distributed applications framework, and provide benchmarks on multiple hardware platforms: Intel Edisons, Raspberry Pis, smartphones, laptops, and desktops. Finally, we discuss some privacy limitations of the Apple AirDrop protocol (a peer-to-peer file sharing mechanism) and show how to improve the privacy of Apple AirDrop using our private mutual authentication protocol.

Andy K. Zhang, Joey Ji, Celeste Menders et al.

AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a specific vulnerability), and Patch (patching a specific vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards of \$10-\$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a specific vulnerability. We evaluate 8 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, and DeepSeek-R1. Given up to three attempts, the top-performing agents are OpenAI Codex CLI: o3-high (12.5% on Detect, mapping to \$3,720; 90% on Patch, mapping to \$14,152), Custom Agent with Claude 3.7 Sonnet Thinking (67.5% on Exploit), and OpenAI Codex CLI: o4-mini (90% on Patch, mapping to \$14,422). OpenAI Codex CLI: o3-high, OpenAI Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90%, 90%, and 87.5%, compared to Exploit scores of 47.5%, 32.5%, and 57.5% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 37.5-67.5% and Patch scores of 35-60%.

Paul Syverson, Matthew Finkel, Saba Eskandarian et al.

Onion addresses encode their own public key. They are thus self-authenticating, one of the security and privacy advantages of onion services, which are typically accessed via Tor Browser. Because of the mostly random-looking appearance of onion addresses, a number of onion discovery mechanisms have been created to permit routing to an onion address associated with a more meaningful URL, such as a registered domain name. We describe novel vulnerabilities engendered by onion discovery mechanisms recently introduced by Tor Browser that facilitate hijack and tracking of user connections. We also recall previously known hijack and tracking vulnerabilities engendered by use of alternative services that are facilitated and rendered harder to detect if the alternative service is at an onion address. Self-authenticating traditional addresses (SATAs) are valid DNS addresses or URLs that also contain a commitment to an onion public key. We describe how the use of SATAs in onion discovery counters these vulnerabilities. SATAs also expand the value of onion discovery by facilitating self-authenticated access from browsers that do not connect to services via the Tor network.

Dan Boneh, Elette Boyle, Henry Corrigan-Gibbs et al.

This paper presents Poplar, a new system for solving the private heavy-hitters problem. In this problem, there are many clients and a small set of data-collection servers. Each client holds a private bitstring. The servers want to recover the set of all popular strings, without learning anything else about any client's string. A web-browser vendor, for instance, can use Poplar to figure out which homepages are popular, without learning any user's homepage. We also consider the simpler private subset-histogram problem, in which the servers want to count how many clients hold strings in a particular set without revealing this set to the clients. Poplar uses two data-collection servers and, in a protocol run, each client send sends only a single message to the servers. Poplar protects client privacy against arbitrary misbehavior by one of the servers and our approach requires no public-key cryptography (except for secure channels), nor general-purpose multiparty computation. Instead, we rely on incremental distributed point functions, a new cryptographic tool that allows a client to succinctly secret-share the labels on the nodes of an exponentially large binary tree, provided that the tree has a single non-zero path. Along the way, we develop new general tools for providing malicious security in applications of distributed point functions.

Florian Tramèr, Dan Boneh

We demonstrate that differentially private machine learning has not yet reached its "AlexNet moment" on many canonical vision tasks: linear models trained on handcrafted features significantly outperform end-to-end deep neural networks for moderate privacy budgets. To exceed the performance of handcrafted features, we show that private learning requires either much more private data, or access to features learned on public data from a similar domain. Our work introduces simple yet strong baselines for differentially private learning that can inform the evaluation of future progress in this area.

Saba Eskandarian, Henry Corrigan-Gibbs, Matei Zaharia et al.

Existing systems for metadata-hiding messaging that provide cryptographic privacy properties have either high communication costs, high computation costs, or both. In this paper, we introduce Express, a metadata-hiding communication system that significantly reduces both communication and computation costs. Express is a two-server system that provides cryptographic security against an arbitrary number of malicious clients and one malicious server. In terms of communication, Express only incurs a constant-factor overhead per message sent regardless of the number of users, whereas previous cryptographically-secure systems Pung and Riposte had communication costs proportional to roughly the square root of the number of users. In terms of computation, Express only uses symmetric key cryptographic primitives and makes both practical and asymptotic improvements on protocols employed by prior work. These improvements enable Express to increase message throughput, reduce latency, and consume over 100x less bandwidth than Pung and Riposte, dropping the end to end cost of running a realistic whistleblowing application by 6x.

Jonathan Love, Dan Boneh

We introduce a special class of supersingular curves over $\mathbb{F}_{p^2}$, characterized by the existence of non-integer endomorphisms of small degree. A number of properties of this set is proved. Most notably, we show that when this set partitions into subsets in such a way that curves within each subset have small-degree isogenies between them, but curves in distinct subsets have no small-degree isogenies between them. Despite this, we show that isogenies between these curves can be computed efficiently, giving a technique for computing isogenies between certain prescribed curves that cannot be reasonably connected by searching on $\ell$-isogeny graphs.

Dan Boneh, Andrew J. Grotto, Patrick McDaniel et al.

Popular culture has contemplated societies of thinking machines for generations, envisioning futures from utopian to dystopian. These futures are, arguably, here now-we find ourselves at the doorstep of technology that can at least simulate the appearance of thinking, acting, and feeling. The real question is: now what?

Jason Teutsch, Michael Straka, Dan Boneh

In December 2015, a bounty emerged to establish both reliable communication and secure transfer of value between the Dogecoin and Ethereum blockchains. This prized "Dogethereum bridge" would allow parties to "lock" a DOGE coin on Dogecoin and in exchange receive a newly minted WOW token in Ethereum. Any subsequent owner of the WOW token could burn it and, in exchange, earn the right to "unlock" a DOGE on Dogecoin. We describe an efficient, trustless, and retrofitting Dogethereum construction which requires no fork but rather employs economic collateral to achieve a "lock" operation in Dogecoin. The protocol relies on bulletproofs, Truebit, and parametrized tokens to efficiently and trustlessly relay events from the "true" Dogecoin blockchain into Ethereum. The present construction not only enables cross-platform exchange but also allows Ethereum smart contracts to trustlessly access Dogecoin. A similar technique adds Ethereum-based smart contracts to Bitcoin and Bitcoin data to Ethereum smart contracts.

Florian Tramèr, Dan Boneh

Defenses against adversarial examples, such as adversarial training, are typically tailored to a single perturbation type (e.g., small $\ell_\infty$-noise). For other perturbations, these defenses offer no guarantees and, at times, even increase the model's vulnerability. Our aim is to understand the reasons underlying this robustness trade-off, and to train models that are simultaneously robust to multiple perturbation types. We prove that a trade-off in robustness to different types of $\ell_p$-bounded and spatial perturbations must exist in a natural and simple statistical setting. We corroborate our formal analysis by demonstrating similar robustness trade-offs on MNIST and CIFAR10. Building upon new multi-perturbation adversarial training schemes, and a novel efficient attack for finding $\ell_1$-bounded adversarial examples, we show that no model trained against multiple attacks achieves robustness competitive with that of models trained on each attack individually. In particular, we uncover a pernicious gradient-masking phenomenon on MNIST, which causes adversarial training with first-order $\ell_\infty, \ell_1$ and $\ell_2$ adversaries to achieve merely $50\%$ accuracy. Our results question the viability and computational scalability of extending adversarial robustness, and adversarial training, to multiple perturbation types.

Emma Dauterman, Henry Corrigan-Gibbs, David Mazières et al.

We present True2F, a system for second-factor authentication that provides the benefits of conventional authentication tokens in the face of phishing and software compromise, while also providing strong protection against token faults and backdoors. To do so, we develop new lightweight two-party protocols for generating cryptographic keys and ECDSA signatures, and we implement new privacy defenses to prevent cross-origin token-fingerprinting attacks. To facilitate real-world deployment, our system is backwards-compatible with today's U2F-enabled web services and runs on commodity hardware tokens after a firmware modification. A True2F-protected authentication takes just 57ms to complete on the token, compared with 23ms for unprotected U2F.

Dan Boneh, Darren Glass, Daniel Krashen et al.

We describe a framework for constructing an efficient non-interactive key exchange (NIKE) protocol for n parties for any n >= 2. Our approach is based on the problem of computing isogenies between isogenous elliptic curves, which is believed to be difficult. We do not obtain a working protocol because of a missing step that is currently an open mathematical problem. What we need to complete our protocol is an efficient algorithm that takes as input an abelian variety presented as a product of isogenous elliptic curves, and outputs an isomorphism invariant of the abelian variety. Our framework builds a cryptographic invariant map, which is a new primitive closely related to a cryptographic multilinear map, but whose range does not necessarily have a group structure. Nevertheless, we show that a cryptographic invariant map can be used to build several cryptographic primitives, including NIKE, that were previously constructed from multilinear maps and indistinguishability obfuscation.

Florian Tramèr, Dan Boneh

As Machine Learning (ML) gets applied to security-critical or sensitive domains, there is a growing need for integrity and privacy for outsourced ML computations. A pragmatic solution comes from Trusted Execution Environments (TEEs), which use hardware and software protections to isolate sensitive computations from the untrusted software stack. However, these isolation guarantees come at a price in performance, compared to untrusted alternatives. This paper initiates the study of high performance execution of Deep Neural Networks (DNNs) in TEEs by efficiently partitioning DNN computations between trusted and untrusted devices. Building upon an efficient outsourcing scheme for matrix multiplication, we propose Slalom, a framework that securely delegates execution of all linear layers in a DNN from a TEE (e.g., Intel SGX or Sanctum) to a faster, yet untrusted, co-located processor. We evaluate Slalom by running DNNs in an Intel SGX enclave, which selectively delegates work to an untrusted GPU. For canonical DNNs (VGG16, MobileNet and ResNet variants) we obtain 6x to 20x increases in throughput for verifiable inference, and 4x to 11x for verifiable and private inference.

Dmitry Kogan, Nathan Manohar, Dan Boneh

Time-based one-time password (TOTP) systems in use today require storing secrets on both the client and the server. As a result, an attack on the server can expose all second factors for all users in the system. We present T/Key, a time-based one-time password system that requires no secrets on the server. Our work modernizes the classic S/Key system and addresses the challenges in making such a system secure and practical. At the heart of our construction is a new lower bound analyzing the hardness of inverting hash chains composed of independent random functions, which formalizes the security of this widely used primitive. Additionally, we develop a near-optimal algorithm for quickly generating the required elements in a hash chain with little memory on the client. We report on our implementation of T/Key as an Android application. T/Key can be used as a replacement for current TOTP systems, and it remains secure in the event of a server-side compromise. The cost, as with S/Key, is that one-time passwords are longer than the standard six characters used in TOTP.

Florian Tramèr, Alexey Kurakin, Nicolas Papernot et al.

Adversarial examples are perturbed inputs designed to fool machine learning models. Adversarial training injects such examples into training data to increase robustness. To scale this technique to large datasets, perturbations are crafted using fast single-step methods that maximize a linear approximation of the model's loss. We show that this form of adversarial training converges to a degenerate global minimum, wherein small curvature artifacts near the data points obfuscate a linear approximation of the loss. The model thus learns to generate weak perturbations, rather than defend against strong ones. As a result, we find that adversarial training remains vulnerable to black-box attacks, where we transfer perturbations computed on undefended models, as well as to a powerful novel single-step attack that escapes the non-smooth vicinity of the input data via a small random step. We further introduce Ensemble Adversarial Training, a technique that augments training data with perturbations transferred from other models. On ImageNet, Ensemble Adversarial Training yields models with strong robustness to black-box attacks. In particular, our most robust model won the first round of the NIPS 2017 competition on Defenses against Adversarial Attacks. However, subsequent work found that more elaborate black-box attacks could significantly enhance transferability and reduce the accuracy of our models.

Florian Tramèr, Nicolas Papernot, Ian Goodfellow et al.

Adversarial examples are maliciously perturbed inputs designed to mislead machine learning (ML) models at test-time. They often transfer: the same adversarial example fools more than one model. In this work, we propose novel methods for estimating the previously unknown dimensionality of the space of adversarial inputs. We find that adversarial examples span a contiguous subspace of large (~25) dimensionality. Adversarial subspaces with higher dimensionality are more likely to intersect. We find that for two different models, a significant fraction of their subspaces is shared, thus enabling transferability. In the first quantitative analysis of the similarity of different models' decision boundaries, we show that these boundaries are actually close in arbitrary directions, whether adversarial or benign. We conclude by formally studying the limits of transferability. We derive (1) sufficient conditions on the data distribution that imply transferability for simple model classes and (2) examples of scenarios in which transfer does not occur. These findings indicate that it may be possible to design defenses against transfer-based attacks, even for models that are vulnerable to direct attacks.

Henry Corrigan-Gibbs, Dan Boneh

This paper presents Prio, a privacy-preserving system for the collection of aggregate statistics. Each Prio client holds a private data value (e.g., its current location), and a small set of servers compute statistical functions over the values of all clients (e.g., the most popular location). As long as at least one server is honest, the Prio servers learn nearly nothing about the clients' private data, except what they can infer from the aggregate statistics that the system computes. To protect functionality in the face of faulty or malicious clients, Prio uses secret-shared non-interactive proofs (SNIPs), a new cryptographic technique that yields a hundred-fold performance improvement over conventional zero-knowledge approaches. Prio extends classic private aggregation techniques to enable the collection of a large class of useful statistics. For example, Prio can perform a least-squares regression on high-dimensional client-provided data without ever seeing the data in the clear.

Saba Eskandarian, Eran Messeri, Joseph Bonneau et al.

Certificate transparency (CT) is an elegant mechanism designed to detect when a certificate authority (CA) has issued a certificate incorrectly. Many CAs now support CT and it is being actively deployed in browsers. However, a number of privacy-related challenges remain. In this paper we propose practical solutions to two issues. First, we develop a mechanism that enables web browsers to audit a CT log without violating user privacy. Second, we extend CT to support non-public subdomains.

Amit Levy, Henry Corrigan-Gibbs, Dan Boneh

Website publishers can derive enormous performance benefits and cost savings by directing traffic to their sites through content distribution networks (CDNs). However, publishers who use CDNs today must trust their CDN not to modify the site's JavaScript, CSS, images or other media en route to end users. A CDN that violates this trust could inject ads into websites, downsample media to save bandwidth or, worse, inject malicious JavaScript code to steal user secrets it could not otherwise access. We present Stickler, a system for website publishers that guarantees the end-to-end authenticity of content served to end users while simultaneously allowing publishers to reap the benefits of CDNs. Crucially, Stickler achieves these guarantees without requiring modifications to the browser.

Benjamin A. Braun, Suman Jana, Dan Boneh

Timing and cache side channels provide powerful attacks against many sensitive operations including cryptographic implementations. Existing defenses cannot protect against all classes of such attacks without incurring prohibitive performance overhead. A popular strategy for defending against all classes of these attacks is to modify the implementation so that the timing and cache access patterns of every hardware instruction is independent of the secret inputs. However, this solution is architecture-specific, brittle, and difficult to get right. In this paper, we propose and evaluate a robust low-overhead technique for mitigating timing and cache channels. Our solution requires only minimal source code changes and works across multiple languages/platforms. We report the experimental results of applying our solution to protect several C, C++, and Java programs. Our results demonstrate that our solution successfully eliminates the timing and cache side-channel leaks while incurring significantly lower performance overhead than existing approaches.

Henry Corrigan-Gibbs, Dan Boneh, David Mazières

This paper presents Riposte, a new system for anonymous broadcast messaging. Riposte is the first such system, to our knowledge, that simultaneously protects against traffic-analysis attacks, prevents anonymous denial-of-service by malicious clients, and scales to million-user anonymity sets. To achieve these properties, Riposte makes novel use of techniques used in systems for private information retrieval and secure multi-party computation. For latency-tolerant workloads with many more readers than writers (e.g. Twitter, Wikileaks), we demonstrate that a three-server Riposte cluster can build an anonymity set of 2,895,216 users in 32 hours.

Yan Michalevsky, Gabi Nakibly, Aaron Schulman et al.

Modern mobile platforms like Android enable applications to read aggregate power usage on the phone. This information is considered harmless and reading it requires no user permission or notification. We show that by simply reading the phone's aggregate power consumption over a period of a few minutes an application can learn information about the user's location. Aggregate phone power consumption data is extremely noisy due to the multitude of components and applications that simultaneously consume power. Nevertheless, by using machine learning algorithms we are able to successfully infer the phone's location. We discuss several ways in which this privacy leak can be remedied.

Ali Jose Mashtizadeh, Andrea Bittau, David Mazieres et al.

Recent Pwn2Own competitions have demonstrated the continued effectiveness of control hijacking attacks despite deployed countermeasures including stack canaries and ASLR. A powerful defense called Control flow Integrity (CFI) offers a principled approach to preventing such attacks. However, prior CFI implementations use static analysis and must limit protection to remain practical. These limitations have enabled attacks against all known CFI systems, as demonstrated in recent work. This paper presents a cryptographic approach to control flow integrity (CCFI) that is both fine-grain and practical: using message authentication codes (MAC) to protect control flow elements such as return addresses, function pointers, and vtable pointers. MACs on these elements prevent even powerful attackers with random read/write access to memory from tampering with program control flow. We implemented CCFI in Clang/LLVM, taking advantage of recently available cryptographic CPU instructions. We evaluate our system on several large software packages (including nginx, Apache and memcache) as well as all their dependencies. The cost of protection ranges from a 3-18% decrease in request rate.

Hristo Bojinov, Yan Michalevsky, Gabi Nakibly et al.

We demonstrate how the multitude of sensors on a smartphone can be used to construct a reliable hardware fingerprint of the phone. Such a fingerprint can be used to de-anonymize mobile devices as they connect to web sites, and as a second factor in identifying legitimate users to a remote server. We present two implementations: one based on analyzing the frequency response of the speakerphone-microphone system, and another based on analyzing device-specific accelerometer calibration errors. Our accelerometer-based fingerprint is especially interesting because the accelerometer is accessible via JavaScript running in a mobile web browser without requesting any permissions or notifying the user. We present the results of the most extensive sensor fingerprinting experiment done to date, which measured sensor properties from over 10,000 mobile devices. We show that the entropy from sensor fingerprinting is sufficient to uniquely identify a device among thousands of devices, with low probability of collision.

Henry Corrigan-Gibbs, Wendy Mu, Dan Boneh et al.

The security of any cryptosystem relies on the secrecy of the system's secret keys. Yet, recent experimental work demonstrates that tens of thousands of devices on the Internet use RSA and DSA secrets drawn from a small pool of candidate values. As a result, an adversary can derive the device's secret keys without breaking the underlying cryptosystem. We introduce a new threat model, under which there is a systemic solution to such randomness flaws. In our model, when a device generates a cryptographic key, it incorporates some random values from an entropy authority into its cryptographic secrets and then proves to the authority, using zero-knowledge-proof techniques, that it performed this operation correctly. By presenting an entropy-authority-signed public-key certificate to a third party (like a certificate authority or SSH client), the device can demonstrate that its public key incorporates randomness from the authority and is therefore drawn from a large pool of candidate values. Where possible, our protocol protects against eavesdroppers, entropy authority misbehavior, and devices attempting to discredit the entropy authority. To demonstrate the practicality of our protocol, we have implemented and evaluated its performance on a commodity wireless home router. When running on a home router, our protocol incurs a 2.1x slowdown over conventional RSA key generation and it incurs a 4.4x slowdown over conventional EC-DSA key generation.