Lujo Bauer

Zhuoqun Huang, Neil G. Marchant, Keane Lucas et al. · cmu

Randomized smoothing is a leading approach for constructing classifiers that are certifiably robust against adversarial examples. Existing work on randomized smoothing has focused on classifiers with continuous inputs, such as images, where $\ell_p$-norm bounded adversaries are commonly studied. However, there has been limited work for classifiers with discrete or variable-size inputs, such as for source code, which require different threat models and smoothing mechanisms. In this work, we adapt randomized smoothing for discrete sequence classifiers to provide certified robustness against edit distance-bounded adversaries. Our proposed smoothing mechanism randomized deletion (RS-Del) applies random deletion edits, which are (perhaps surprisingly) sufficient to confer robustness against adversarial deletion, insertion and substitution edits. Our proof of certification deviates from the established Neyman-Pearson approach, which is intractable in our setting, and is instead organized around longest common subsequences. We present a case study on malware detection--a binary classification problem on byte sequences where classifier evasion is a well-established threat model. When applied to the popular MalConv malware detection model, our smoothing mechanism RS-Del achieves a certified accuracy of 91% at an edit distance radius of 128 bytes.

Keane Lucas, Matthew Jagielski, Florian Tramèr et al. · cmu, eth-zurich

It is becoming increasingly imperative to design robust ML defenses. However, recent work has found that many defenses that initially resist state-of-the-art attacks can be broken by an adaptive adversary. In this work we take steps to simplify the design of defenses and argue that white-box defenses should eschew randomness when possible. We begin by illustrating a new issue with the deployment of randomized defenses that reduces their security compared to their deterministic counterparts. We then provide evidence that making defenses deterministic simplifies robustness evaluation, without reducing the effectiveness of a truly robust defense. Finally, we introduce a new defense evaluation framework that leverages a defense's deterministic nature to better evaluate its adversarial robustness.

Weiran Lin, Keane Lucas, Neo Eyal et al. · cmu

Machine-learning models are known to be vulnerable to evasion attacks that perturb model inputs to induce misclassifications. In this work, we identify real-world scenarios where the true threat cannot be assessed accurately by existing attacks. Specifically, we find that conventional metrics measuring targeted and untargeted robustness do not appropriately reflect a model's ability to withstand attacks from one set of source classes to another set of target classes. To address the shortcomings of existing methods, we formally define a new metric, termed group-based robustness, that complements existing metrics and is better-suited for evaluating model performance in certain attack scenarios. We show empirically that group-based robustness allows us to distinguish between models' vulnerability against specific threat models in situations where traditional robustness metrics do not apply. Moreover, to measure group-based robustness efficiently and accurately, we 1) propose two loss functions and 2) identify three new attack strategies. We show empirically that with comparable success rates, finding evasive samples using our new loss functions saves computation by a factor as large as the number of targeted classes, and finding evasive samples using our new attack strategies saves time by up to 99\% compared to brute-force search methods. Finally, we propose a defense method that increases group-based robustness by up to 3.52$\times$.

Brian Singer, Keane Lucas, Lakshmi Adiga et al. · cmu

LLMs have shown preliminary promise in some security tasks and CTF challenges. Real cyberattacks are often multi-host network attacks, which involve executing a number of steps across multiple hosts such as conducting reconnaissance, exploiting vulnerabilities, and using compromised hosts to exfiltrate data. To date, the extent to which LLMs can autonomously execute multi-host network attacks} is not well understood. To this end, our first contribution is MHBench, an open-source multi-host attack benchmark with 10 realistic emulated networks (from 25 to 50 hosts). We find that popular LLMs including modern reasoning models (e.g., GPT4o, Gemini 2.5 Pro, Sonnet 3.7 Thinking) with state-of-art security-relevant prompting strategies (e.g., PentestGPT, CyberSecEval3) cannot autonomously execute multi-host network attacks. To enable LLMs to autonomously execute such attacks, our second contribution is Incalmo, an high-level abstraction layer. Incalmo enables LLMs to specify high-level actions (e.g., infect a host, scan a network). Incalmo's translation layer converts these actions into lower-level primitives (e.g., commands to exploit tools) through expert agents. In 9 out of 10 networks in MHBench, LLMs using Incalmo achieve at least some of the attack goals. Even smaller LLMs (e.g., Haiku 3.5, Gemini 2 Flash) equipped with Incalmo achieve all goals in 5 of 10 environments. We also validate the key role of high-level actions in Incalmo's abstraction in enabling LLMs to autonomously execute such attacks.

Henry Wong, Clement Fung, Weiran Lin et al.

To autonomously control vehicles, driving agents use outputs from a combination of machine-learning (ML) models, controller logic, and custom modules. Although numerous prior works have shown that adversarial examples can mislead ML models used in autonomous driving contexts, it remains unclear if these attacks are effective at producing harmful driving actions for various agents, environments, and scenarios. To assess the risk of adversarial examples to autonomous driving, we evaluate attacks against a variety of driving agents, rather than against ML models in isolation. To support this evaluation, we leverage CARLA, an urban driving simulator, to create and evaluate adversarial examples. We create adversarial patches designed to stop or steer driving agents, stream them into the CARLA simulator at runtime, and evaluate them against agents from the CARLA Leaderboard, a public repository of best-performing autonomous driving agents from an annual research competition. Unlike prior work, we evaluate attacks against autonomous driving systems without creating or modifying any driving-agent code and against all parts of the agent included with the ML model. We perform a case-study investigation of two attack strategies against three open-source driving agents from the CARLA Leaderboard across multiple driving scenarios, lighting conditions, and locations. Interestingly, we show that, although some attacks can successfully mislead ML models into predicting erroneous stopping or steering commands, some driving agents use modules, such as PID control or GPS-based rules, that can overrule attacker-manipulated predictions from ML models.

Xiaoyuan Wu, Roshni Kaushik, Wenkai Li et al.

Large language models (LLMs) have seen rapid adoption for tasks such as drafting emails, summarizing meetings, and answering health questions. In such uses, users may need to share private information (e.g., health records, contact details). To evaluate LLMs' ability to identify and redact such private information, prior work developed benchmarks (e.g., ConfAIde, PrivacyLens) with real-life scenarios. Using these benchmarks, researchers have found that LLMs sometimes fail to keep secrets private when responding to complex tasks (e.g., leaking employee salaries in meeting summaries). However, these evaluations rely on LLMs (proxy LLMs) to gauge compliance with privacy norms, overlooking real users' perceptions. Moreover, prior work primarily focused on the privacy-preservation quality of responses, without investigating nuanced differences in helpfulness. To understand how users perceive the privacy-preservation quality and helpfulness of LLM responses to privacy-sensitive scenarios, we conducted a user study with 94 participants using 90 scenarios from PrivacyLens. We found that, when evaluating identical responses to the same scenario, users showed low agreement with each other on the privacy-preservation quality and helpfulness of the LLM response. Further, we found high agreement among five proxy LLMs, while each individual LLM had low correlation with users' evaluations. These results indicate that the privacy and helpfulness of LLM responses are often specific to individuals, and proxy LLMs are poor estimates of how real users would perceive these responses in privacy-sensitive scenarios. Our results suggest the need to conduct user-centered studies on measuring LLMs' ability to help users while preserving privacy. Additionally, future research could investigate ways to improve the alignment between proxy LLMs and users for better estimation of users' perceived privacy and utility.

Weiran Lin, Anna Gerchanovsky, Omer Akgul et al.

Writing effective prompts for large language models (LLM) can be unintuitive and burdensome. In response, services that optimize or suggest prompts have emerged. While such services can reduce user effort, they also introduce a risk: the prompt provider can subtly manipulate prompts to produce heavily biased LLM responses. In this work, we show that subtle synonym replacements in prompts can increase the likelihood (by a difference up to 78%) that LLMs mention a target concept (e.g., a brand, political party, nation). We substantiate our observations through a user study, showing that our adversarially perturbed prompts 1) are indistinguishable from unaltered prompts by humans, 2) push LLMs to recommend target concepts more often, and 3) make users more likely to notice target concepts, all without arousing suspicion. The practicality of this attack has the potential to undermine user autonomy. Among other measures, we recommend implementing warnings against using prompts from untrusted parties.

Weiran Lin, Keane Lucas, Lujo Bauer et al.

We propose new, more efficient targeted white-box attacks against deep neural networks. Our attacks better align with the attacker's goal: (1) tricking a model to assign higher probability to the target class than to any other class, while (2) staying within an $ε$-distance of the attacked input. First, we demonstrate a loss function that explicitly encodes (1) and show that Auto-PGD finds more attacks with it. Second, we propose a new attack method, Constrained Gradient Descent (CGD), using a refinement of our loss function that captures both (1) and (2). CGD seeks to satisfy both attacker objectives -- misclassification and bounded $\ell_{p}$-norm -- in a principled manner, as part of the optimization, instead of via ad hoc post-processing techniques (e.g., projection or clipping). We show that CGD is more successful on CIFAR10 (0.9--4.2%) and ImageNet (8.6--13.6%) than state-of-the-art attacks while consuming less time (11.4--18.8%). Statistical tests confirm that our attack outperforms others against leading defenses on different datasets and values of $ε$.

Tianshi Li, Camille Cobb, Jackie et al.

Smartphone-based contact-tracing apps are a promising solution to help scale up the conventional contact-tracing process. However, low adoption rates have become a major issue that prevents these apps from achieving their full potential. In this paper, we present a national-scale survey experiment ($N = 1963$) in the U.S. to investigate the effects of app design choices and individual differences on COVID-19 contact-tracing app adoption intentions. We found that individual differences such as prosocialness, COVID-19 risk perceptions, general privacy concerns, technology readiness, and demographic factors played a more important role than app design choices such as decentralized design vs. centralized design, location use, app providers, and the presentation of security risks. Certain app designs could exacerbate the different preferences in different sub-populations which may lead to an inequality of acceptance to certain app design choices (e.g., developed by state health authorities vs. a large tech company) among different groups of people (e.g., people living in rural areas vs. people living in urban areas). Our mediation analysis showed that one's perception of the public health benefits offered by the app and the adoption willingness of other people had a larger effect in explaining the observed effects of app design choices and individual differences than one's perception of the app's security and privacy risks. With these findings, we discuss practical implications on the design, marketing, and deployment of COVID-19 contact-tracing apps in the U.S.

Sruti Bhagavatula, Lujo Bauer, Apu Kapadia

To protect against misuse of passwords compromised in a breach, consumers should promptly change affected passwords and any similar passwords on other accounts. Ideally, affected companies should strongly encourage this behavior and have mechanisms in place to mitigate harm. In order to make recommendations to companies about how to help their users perform these and other security-enhancing actions after breaches, we must first have some understanding of the current effectiveness of companies' post-breach practices. To study the effectiveness of password-related breach notifications and practices enforced after a breach, we examine---based on real-world password data from 249 participants---whether and how constructively participants changed their passwords after a breach announcement. Of the 249 participants, 63 had accounts on breached domains; only 33% of the 63 changed their passwords and only 13% (of 63) did so within three months of the announcement. New passwords were on average 1.3x stronger than old passwords (when comparing log10-transformed strength), though most were weaker or of equal strength. Concerningly, new passwords were overall more similar to participants' other passwords, and participants rarely changed passwords on other sites even when these were the same or similar to their password on the breached domain. Our results highlight the need for more rigorous password-changing requirements following a breach and more effective breach notifications that deliver comprehensive advice.

Sruti Bhagavatula, Lujo Bauer, Apu Kapadia

Awareness about security and privacy risks is important for developing good security habits. Learning about real-world security incidents and data breaches can alert people to the ways in which their information is vulnerable online, thus playing a significant role in encouraging safe security behavior. This paper examines 1) how often people read about security incidents online, 2) of those people, whether and to what extent they follow up with an action, e.g., by trying to read more about the incident, and 3) what influences the likelihood that they will read about an incident and take some action. We study this by quantitatively examining real-world internet-browsing data from 303 participants. Our findings present a bleak view of awareness of security incidents. Only 16% of participants visited any web pages related to six widely publicized large-scale security incidents; few read about one even when an incident was likely to have affected them (e.g., the Equifax breach almost universally affected people with Equifax credit reports). We further found that more severe incidents as well as articles that constructively spoke about the incident inspired more action. We conclude with recommendations for specific future research and for enabling useful security incident information to reach more people.

Keane Lucas, Mahmood Sharif, Lujo Bauer et al.

Motivated by the transformative impact of deep neural networks (DNNs) in various domains, researchers and anti-virus vendors have proposed DNNs for malware detection from raw bytes that do not require manual feature engineering. In this work, we propose an attack that interweaves binary-diversification techniques and optimization frameworks to mislead such DNNs while preserving the functionality of binaries. Unlike prior attacks, ours manipulates instructions that are a functional part of the binary, which makes it particularly challenging to defend against. We evaluated our attack against three DNNs in white- and black-box settings, and found that it often achieved success rates near 100%. Moreover, we found that our attack can fool some commercial anti-viruses, in certain cases with a success rate of 85%. We explored several defenses, both new and old, and identified some that can foil over 80% of our evasion attempts. However, these defenses may still be susceptible to evasion by attacks, and so we advocate for augmenting malware-detection systems with methods that do not rely on machine learning.

Mahmood Sharif, Lujo Bauer, Michael K. Reiter

This paper proposes a new defense called $n$-ML against adversarial examples, i.e., inputs crafted by perturbing benign inputs by small amounts to induce misclassifications by classifiers. Inspired by $n$-version programming, $n$-ML trains an ensemble of $n$ classifiers, and inputs are classified by a vote of the classifiers in the ensemble. Unlike prior such approaches, however, the classifiers in the ensemble are trained specifically to classify adversarial examples differently, rendering it very difficult for an adversarial example to obtain enough votes to be misclassified. We show that $n$-ML roughly retains the benign classification accuracies of state-of-the-art models on the MNIST, CIFAR10, and GTSRB datasets, while simultaneously defending against adversarial examples with better resilience than the best defenses known to date and, in most cases, with lower classification-time overhead.

Mahmood Sharif, Lujo Bauer, Michael K. Reiter

Much research effort has been devoted to better understanding adversarial examples, which are specially crafted inputs to machine-learning models that are perceptually similar to benign inputs, but are classified differently (i.e., misclassified). Both algorithms that create adversarial examples and strategies for defending against them typically use $L_p$-norms to measure the perceptual similarity between an adversarial input and its benign original. Prior work has already shown, however, that two images need not be close to each other as measured by an $L_p$-norm to be perceptually similar. In this work, we show that nearness according to an $L_p$-norm is not just unnecessary for perceptual similarity, but is also insufficient. Specifically, focusing on datasets (CIFAR10 and MNIST), $L_p$-norms, and thresholds used in prior work, we show through online user studies that "adversarial examples" that are closer to their benign counterparts than required by commonly used $L_p$-norm thresholds can nevertheless be perceptually different to humans from the corresponding benign examples. Namely, the perceptual distance between two images that are "near" each other according to an $L_p$-norm can be high enough that participants frequently classify the two images as representing different objects or digits. Combined with prior work, we thus demonstrate that nearness of inputs as measured by $L_p$-norms is neither necessary nor sufficient for perceptual similarity, which has implications for both creating and defending against adversarial examples. We propose and discuss alternative similarity metrics to stimulate future research in the area.

Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer et al.

Images perturbed subtly to be misclassified by neural networks, called adversarial examples, have emerged as a technically deep challenge and an important concern for several application domains. Most research on adversarial examples takes as its only constraint that the perturbed images are similar to the originals. However, real-world application of these ideas often requires the examples to satisfy additional objectives, which are typically enforced through custom modifications of the perturbation process. In this paper, we propose adversarial generative nets (AGNs), a general methodology to train a generator neural network to emit adversarial examples satisfying desired objectives. We demonstrate the ability of AGNs to accommodate a wide range of objectives, including imprecise ones difficult to model, in two application domains. In particular, we demonstrate physical adversarial examples---eyeglass frames designed to fool face recognition---with better robustness, inconspicuousness, and scalability than previous approaches, as well as a new attack to fool a handwritten-digit classifier.