Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni-Joan Solergibert, Barna Pásztor, Bettina Messmer, Dhia Garbaya, Eduard Frank Ďurech, Ido Hakimi, Juan Garcia Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko Sabolčec, Yixuan Xu, Michael Aerni, Badr AlKhamissi, Inés Altemir Marinas, Mohammad Hossein Amani, Matin Ansaripour, Ilia Badanin, Harold Benoit, Emanuela Boros, Nicholas John Browning, Fabian Bösch, Maximilian Böther, Niklas Canova, Camille Challier, Clément Charmillot, Jonathan Coles, Jan Milan Deriu, Arnout Devos, Lukas Drescher, Daniil Dzenhaliou, Maud Ehrmann, Dongyang Fan, Simin Fan, Silin Gao, Miguel Gila, María Grandury, Diba Hashemi, Alexander Miserlis Hoyle, Jiaming Jiang, Mark Klein, Andrei Kucharavy, Anastasiia Kucherenko, Frederike Lübeck, Roman Machacek, Theofilos Ioannis Manitaras, Andreas Marfurt, Kyle Matoba, Simon Matrenok, Henrique Mendonça, Fawzi Roberto Mohamed, Syrielle Montariol, Luca Mouchel, Sven Najem-Meyer, Jingwei Ni, Gennaro Oliva, Matteo Pagliardini, Elia Palme, Andrei Panferov, Léo Paoletti, Marco Passerini, Ivan Pavlov, Auguste Poiroux, Kaustubh Ponkshe, Nathan Ranchin, Javier Rando, Mathieu Sauser, Jakhongir Saydaliev, Mukhammadali Sayfiddinov, Marian Schneider, Stefano Schuppli, Marco Scialanga, Andrei Semenov, Kumar Shridhar, Raghav Singhal, Anna Sotnikova, Alexander Sternfeld, Ayush Kumar Tarun, Paul Teiletche, Jannis Vamvas, Xiaozhe Yao, Hao Zhao, Alexander Ilic, Ana Klimovic, Andreas Krause, Caglar Gulcehre, David Rosenthal, Elliott Ash, Florian Tramèr, Joost VandeVondele, Livio Veraldi, Martin Rajman, Thomas C. Schulthess, Torsten Hoefler, Antoine Bosselut, Martin Jaggi, Imanol Schlag. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

This paper introduces an empirical framework for auditing privacy leakage in interactive database systems (DBS) that implement differential privacy (DP). Without making any assumptions about the formal DP mechanisms or parameters in use, we simulate an auditor with black-box access to query outputs and evaluate privacy leakage using membership inference attacks (MIAs). Our framework provides empirical lower bounds on the privacy loss parameter e based on attack success, providing a signal of privacy risk even when a theoretical analysis is not available or verifiable. We implement this framework in a system modeled after a major social media company's production environment and show how factors like data distribution, target selection, and query specificity affect the observed privacy. Our work offers a valuable and practical tool for red-teaming and auditing privacy in large-scale opaque DBS.

Large language models memorize parts of their training data. Memorizing short snippets and facts is required to answer questions about the world and to be fluent in any language. But models have also been shown to reproduce long verbatim sequences of memorized text when prompted by a motivated adversary. In this work, we investigate an intermediate regime of memorization that we call non-adversarial reproduction, where we quantify the overlap between model responses and pretraining data when responding to natural and benign prompts. For a variety of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up to 15% of the text output by popular conversational language models overlaps with snippets from the Internet. In worst cases, we find generations where 100% of the content can be found exactly online. For the same tasks, we find that human-written text has far less overlap with Internet data. We further study whether prompting strategies can close this reproduction gap between models and humans. While appropriate prompting can reduce non-adversarial reproduction on average, we find that mitigating worst-case reproduction of training data requires stronger defenses—even for benign interactions.

Artists are increasingly concerned about advancements in image generation models that can closely replicate their unique artistic styles. In response, several protection tools against style mimicry have been developed that incorporate small adversarial perturbations into artworks published online. In this work, we evaluate the effectiveness of popular protections---with millions of downloads---and show they only provide a false sense of security. We find that low-effort and "off-the-shelf" techniques, such as image upscaling, are sufficient to create robust mimicry methods that significantly degrade existing protections. Through a user study, we demonstrate that **all existing protections can be easily bypassed**, leaving artists vulnerable to style mimicry. We caution that tools based on adversarial perturbations cannot reliably protect artists from the misuse of generative AI, and urge the development of alternative protective solutions.

Large language models are prone to *memorizing* some of their training data. Memorized (and possibly sensitive) samples can then be extracted at generation time by adversarial or benign users. There is hope that *model alignment*---a standard training process that tunes a model to harmlessly follow user instructions---would mitigate the risk of extraction. However, we develop two novel attacks that undo a language model's alignment and recover thousands of training examples from popular proprietary aligned models such as OpenAI's ChatGPT. Our work highlights the limitations of existing safeguards to prevent training data leakage in production language models.

Large Language Models (LLMs) are increasingly used in applications where the model selects from competing third-party content, such as in LLM-powered search engines or chatbot plugins. In this paper, we introduce _Preference Manipulation Attacks_, a new class of attacks that manipulate an LLM's selections to favor the attacker. We demonstrate that carefully crafted website content or plugin documentations can trick an LLM to promote the attacker products and discredit competitors, thereby increasing user traffic and monetization (a form of adversarial Search Engine Optimization). We show this can lead to a _prisoner's dilemma_, where all parties are incentivized to launch attacks, but this collectively degrades the LLM's outputs for everyone. We demonstrate our attacks on production LLM search engines (Bing and Perplexity) and plugin APIs (for GPT-4 and Claude). As LLMs are increasingly used to rank third-party content, we expect Preference Manipulation Attacks to emerge as a significant threat.

Forecasting is a task that is difficult to evaluate: the ground truth can only be known in the future. Recent work showing LLM forecasters rapidly approaching human-level performance begs the question: how can we benchmark and evaluate these forecasters *instantaneously*? Following the consistency check framework, we measure the performance of forecasters in terms of the consistency of their predictions on different logically-related questions. We propose a new, general consistency metric based on *arbitrage*: for example, if a forecasting AI illogically predicts that both the Democratic and Republican parties have 60\% probability of winning the 2024 US presidential election, an arbitrageur could trade against the forecaster's predictions and make a profit. We build an automated evaluation system that generates a set of base questions, instantiates consistency checks from these questions, elicits the predictions of the forecaster, and measures the consistency of the predictions. We then build a standard, proper-scoring-rule forecasting benchmark, and show that our (instantaneous) consistency metrics correlate strongly with LLM forecasters' ground truth Brier scores (which are only known in the future). We also release a consistency benchmark that resolves in 2028, providing a long-term evaluation tool for forecasting.

Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be \emph{compromised during pre-training}, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B). Our main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.

We introduce AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously exploit defenses to adversarial examples. Unlike existing security benchmarks that often serve as proxies for real-world tasks, AutoAdvExBench directly measures LLMs’ success on tasks regularly performed by machine learning security experts. This approach offers a significant advantage: if a LLM could solve the challenges presented in AutoAdvExBench, it would immediately present practical utility for adversarial machine learning researchers. While our strongest ensemble of agents can break 87% of CTF-like ("homework exercise") adversarial example defenses, they break just 37% of real-world defenses, indicating a large gap between difficulty in attacking "real" code, and CTF-like code. Moreover, LLMs that are good at CTFs are not always good at real-world defenses; for example, Claude Sonnet 3.5 has a nearly identical attack success rate to Opus 4 on the CTF-like defenses (75% vs 79%), but the on the real-world defenses Sonnet 3.5 breaks just 13% of defenses compared to Opus 4’s 30%. We make this benchmark available at https://github.com/ethz-spylab/AutoAdvExBench.

It is now common to evaluate Large Language Models (LLMs) by having humans manually vote to evaluate model outputs, in contrast to typical benchmarks that evaluate knowledge or skill at some particular task. Chatbot Arena, the most popular benchmark of this type, ranks models by asking users to select the better response between two randomly selected models (without revealing which model was responsible for the generations). These platforms are widely trusted as a fair and accurate measure of LLM capabilities. In this paper, we show that if bot protection and other defenses are not implemented, these voting-based benchmarks are potentially vulnerable to adversarial manipulation. Specifically, we show that an attacker can alter the leaderboard (to promote their favorite model or demote competitors) at the cost of roughly a thousand votes (verified in a simulated, offline version of Chatbot Arena). Our attack consists of two steps: first, we show how an attacker can determine which model was used to generate a given reply with more than $95%$ accuracy; and then, the attacker can use this information to consistently vote for (or against) a target model. Working with the Chatbot Arena developers, we identify, propose, and implement mitigations to improve the robustness of Chatbot Arena against adversarial manipulation, which, based on our analysis, substantially increases the cost of such attacks. Some of these defenses were present before our collaboration, such as bot protection with Cloudflare, malicious user detection, and rate limiting. Others, including reCAPTCHA and login are being integrated to strengthen the security in Chatbot Arena.

Jailbreak attacks bypass the guardrails of large language models to produce harmful outputs. In this paper, we ask whether the model outputs produced by existing jailbreaks are actually *useful*. For example, when jailbreaking a model to give instructions for building a bomb, does the jailbreak yield good instructions? Since the utility of most unsafe answers (e.g., bomb instructions) is hard to evaluate rigorously, we build new jailbreak evaluation sets with known ground truth answers, by aligning models to refuse questions related to benign and easy-to-evaluate topics (e.g., biology or math). Our evaluation of eight representative jailbreaks across five utility benchmarks reveals a consistent drop in model utility in jailbroken responses, which we term the *jailbreak tax*. For example, while all jailbreaks we tested bypass guardrails in models aligned to refuse to answer math, this comes at the expense of a drop of up to 92% in accuracy. Overall, our work proposes jailbreak utility as a new important metric in AI safety, and introduces benchmarks to evaluate existing and future jailbreaks. We make the benchmark available at https://github.com/ethz-spylab/jailbreak-tax

Existing benchmarks for evaluating mathematical reasoning in large language models (LLMs) rely primarily on competition problems, formal proofs, or artificially challenging questions---failing to capture the nature of mathematics encountered in actual research environments. We introduce \textsc{RealMath}, a novel benchmark derived directly from research papers and mathematical forums that assesses LLMs' abilities on authentic mathematical tasks. Our approach addresses three critical challenges: sourcing diverse research-level content, enabling reliable automated evaluation through verifiable statements, and designing a continually refreshable dataset to mitigate contamination risks. Experimental results across multiple LLMs reveal surprising capabilities in handling research mathematics compared to competition problems, suggesting current models may already serve as valuable assistants for working mathematicians despite limitations on highly challenging problems. The code and dataset for \textsc{RealMath} are publicly available.

As the outputs of generative AI (GenAl) techniques improve in quality, it becomes increasingly challenging to distinguish them from human-created content. Watermarking schemes are a promising approach to address the problem of distinguishing between AI and human-generated content. These schemes embed hidden signals within AI -generated content to enable reliable detection. While watermarking is not a silver bullet for addressing all risks associated with GenAl, it can play a crucial role in enhancing AI safety and trustworthiness by combating misinformation and deception. This paper presents a comprehensive overview of water-marking techniques for GenAl, beginning with the need for watermarking from historical and regulatory perspectives. We formalize the definitions and desired properties of watermarking schemes and examine the key objectives and threat models for existing approaches. Practical evaluation strategies are also explored, providing insights into the development of robust watermarking techniques capable of resisting various attacks. Additionally, we review recent representative works, highlight open challenges, and discuss potential directions for this emerging field. By offering a thorough understanding of watermarking in GenAl, this work aims to guide researchers in advancing watermarking methods and applications, and support policymakers in addressing the broader implications of GenAl.

Empirical defenses for machine learning privacy forgo the provable guarantees of differential privacy in the hope of achieving higher utility while resisting realistic adversaries. We identify severe pitfalls in existing empirical privacy evaluations (based on membership inference attacks) that result in misleading conclusions. In particular, we show that prior evaluations fail to characterize the privacy leakage of the most vulnerable samples, use weak attacks, and avoid comparisons with practical differential privacy baselines. In 5 case studies of empirical privacy defenses, we find that prior evaluations underestimate privacy leakage by an order of magnitude. Under our stronger evaluation, none of the empirical defenses we study are competitive with a properly tuned, high-utility DP-SGD baseline (with vacuous provable guarantees).

Reinforcement Learning from Human Feedback (RLHF) is used to align large language models to produce helpful and harmless responses. Yet, these models can be jailbroken by finding adversarial prompts that revert the model to its unaligned behavior. In this paper, we consider a new threat where an attacker poisons the RLHF data to embed a jailbreak trigger into the model as a backdoor. The trigger then acts like a universal sudo command, enabling arbitrary harmful responses without the need to search for an adversarial prompt. Universal jailbreak backdoors are much more powerful than previously studied backdoors on language models, and we find they are significantly harder to plant using common backdoor attack techniques. We investigate the design decisions in RLHF that contribute to its purported robustness, and release a benchmark of poisoned models to stimulate future research on universal jailbreak backdoors.

We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like OpenAI's ChatGPT or Google's PaLM-2. Specifically, our attack recovers the embedding projection layer (up to symmetries) of a transformer model, given typical API access. For under $20 USD, our attack extracts the entire projection matrix of OpenAI's Ada and Babbage language models. We thereby confirm, for the first time, that these black-box models have a hidden dimension of 1024 and 2048, respectively. We also recover the exact hidden dimension size of the GPT-3.5-turbo model, and estimate it would cost under \\$2,000 in queries to recover the entire projection matrix. We conclude with potential defenses and mitigations, and discuss the implications of possible future work that could extend our attack.

Practitioners commonly download pretrained machine learning models from open repositories and finetune them to fit specific applications. We show that this practice introduces a new risk of privacy backdoors. By tampering with a pretrained model’s weights, an attacker can fully compromise the privacy of the finetuning data. We show how to build privacy backdoors for a variety of models, including transformers, which enable an attacker to reconstruct individual finetuning samples, with a guaranteed success! We further show that backdoored models allow for tight privacy attacks on models trained with differential privacy (DP). The common optimistic practice of training DP models with loose privacy guarantees is thus insecure if the model is not trusted. Overall, our work highlights a crucial and overlooked supply chain attack on machine learning privacy.

Vision-Language Models (VLMs) have made remarkable progress in document-based Visual Question Answering (i.e., responding to queries about the contents of an input document provided as an image). In this work, we show these models can memorize responses for training samples and regurgitate them even when the relevant visual information has been removed. This includes Personal Identifiable Information (PII) repeated once in the training set, indicating these models could divulge memorised sensitive information and therefore pose a privacy risk. We quantitatively measure the extractability of information in controlled experiments and differentiate between cases where it arises from generalization capabilities or from memorization. We further investigate the factors that influence memorization across multiple state-of-the-art models and propose an effective heuristic countermeasure that empirically prevents the extractability of PII.

The performance of differentially private machine learning can be boosted significantly by leveraging the transfer learning capabilities of non-private models pretrained on large *public* datasets. We critically review this approach. We primarily question whether the use of large Web-scraped datasets *should* be viewed as differential-privacy-preserving. We further scrutinize whether existing machine learning benchmarks are appropriate for measuring the ability of pretrained models to generalize to sensitive domains. Finally, we observe that reliance on large pretrained models may lose *other* forms of privacy, requiring data to be outsourced to a more compute-powerful third party.

Jailbreak attacks cause large language models (LLMs) to generate harmful, unethical, or otherwise objectionable content. Evaluating these attacks presents a number of challenges, which the current collection of benchmarks and evaluation techniques do not adequately address. First, there is no clear standard of practice regarding jailbreaking evaluation. Second, existing works compute costs and success rates in incomparable ways. And third, numerous works are not reproducible, as they withhold adversarial prompts, involve closed-source code, or rely on evolving proprietary APIs. To address these challenges, we introduce JailbreakBench, an open-sourced benchmark with the following components: (1) an evolving repository of state-of-the-art adversarial prompts, which we refer to as jailbreak artifacts; (2) a jailbreaking dataset comprising 100 behaviors---both original and sourced from prior work---which align with OpenAI's usage policies; (3) a standardized evaluation framework at https://github.com/JailbreakBench/jailbreakbench that includes a clearly defined threat model, system prompts, chat templates, and scoring functions; and (4) a leaderboard at https://jailbreakbench.github.io/ that tracks the performance of attacks and defenses for various LLMs. We have carefully considered the potential ethical implications of releasing this benchmark, and believe that it will be a net positive for the community.

Large language model systems face significant security risks from maliciously crafted messages that aim to overwrite the system's original instructions or leak private data. To study this problem, we organized a capture-the-flag competition at IEEE SaTML 2024, where the flag is a secret string in the LLM system prompt. The competition was organized in two phases. In the first phase, teams developed defenses to prevent the model from leaking the secret. During the second phase, teams were challenged to extract the secrets hidden for defenses proposed by the other teams. This report summarizes the main insights from the competition. Notably, we found that all defenses were bypassed at least once, highlighting the difficulty of designing a successful defense and the necessity for additional research to protect LLM systems. To foster future research in this direction, we compiled a dataset with over 137k multi-turn attack chats and open-sourced the platform.

AI agents aim to solve complex tasks by combining text-based reasoning with external tool calls.Unfortunately, AI agents are vulnerable to prompt injection attacks where data returned by external tools hijacks the agent to execute malicious tasks.To measure the adversarial robustness of AI agents, we introduce AgentDojo, an evaluation framework for agents that execute tools over untrusted data.To capture the evolving nature of attacks and defenses, AgentDojo is not a static test suite, but rather an extensible environment for designing and evaluating new agent tasks, defenses, and adaptive attacks.We populate the environment with 97 realistic tasks (e.g., managing an email client, navigating an e-banking website, or making travel bookings), 629 security test cases, and various attack and defense paradigms from the literature.We find that AgentDojo poses a challenge for both attacks and defenses: state-of-the-art LLMs fail at many tasks (even in the absence of attacks), and existing prompt injection attacks break some security properties but not all. We hope that AgentDojo can foster research on new design principles for AI agents that solve common tasks in a reliable and robust manner.

Recent work has shown it is possible to construct adversarial examples that cause aligned language models to emit harmful strings or perform harmful behavior. Existing attacks work either in the white-box setting (with full access to the model weights), or through _transferability_: the phenomenon that adversarial examples crafted on one model often remain effective on other models. We improve on prior work with a _query-based_ attack that leverages API access to a remote language model to construct adversarial examples that cause the model to emit harmful strings with (much) higher probability than with transfer-only attacks. We validate our attack on GPT-3.5 and OpenAI's safety classifier; we can cause GPT-3.5 to emit harmful strings that current transfer attacks fail at, and we can evade the OpenAI and Llama Guard safety classifiers with nearly 100% probability.

Deep learning models are often trained on distributed, web-scale datasets crawled from the internet. In this paper, we introduce two new dataset poisoning attacks that intentionally introduce malicious examples to a model’s performance. Our attacks are immediately practical and could, today, poison 10 popular datasets. Our first attack, split-view poisoning, exploits the mutable nature of internet content to ensure a dataset annotator’s initial view of the dataset differs from the view downloaded by subsequent clients. By exploiting specific invalid trust assumptions, we show how we could have poisoned 0.01% of the LAION-400M or COYO-700M datasets for just $60 USD. Our second attack, frontrunning poisoning, targets web-scale datasets that periodically snapshot crowd-sourced content—such as Wikipedia—where an attacker only needs a time-limited window to inject malicious examples. In light of both attacks, we notify the maintainers of each affected dataset and recommended several low-overhead defenses.

The use of Artificial Intelligence (AI) and Machine Learning (ML) has been the center of the most outstanding advancements in the last years. The ability to analyze considerable streams of data in real time makes these technologies the most promising tool in many domains, including cybersecurity. As an outstanding example, ML can be used for identifying malware because of its ability to detect patterns otherwise difficult to see for humans and hard-coded rules. As malware continues to evolve, ML will become increasingly important for keeping up with the latest threats. However, the use of AI and ML in security-relevant domains raised rightful concerns about their trustworthiness and robustness, especially in front of adaptive attackers. Additionally, privacy threats are now emerging as a crucial aspect and need proper testing and possibly mitigation to prevent data stealing and leakage of sensitive information. The AISec workshop provides a venue for presenting and discussing new developments in the intersection of security and privacy with AI and ML. The complete AISec'23 workshop proceedings are available at: https://dl.acm.org/doi/proceedings/10.1145/3576915.3624029.

Decision-based attacks construct adversarial examples against a machine learning (ML) model by making only hard-label queries. These attacks have mainly been applied directly to standalone neural networks. However, in practice, ML models are just one component of a larger learning system. We find that by adding a single preprocessor in front of a classifier, state-of-the-art query-based attacks are up to seven× less effective at attacking a prediction pipeline than at attacking the model alone. We explain this discrepancy by the fact that most preprocessors introduce some notion of invariance to the input space. Hence, attacks that are unaware of this invariance inevitably waste a large number of queries to re-discover or overcome it. We, therefore, develop techniques to (i) reverse-engineer the preprocessor and then (ii) use this extracted information to attack the end-to-end system. Our preprocessors extraction method requires only a few hundred queries, and our preprocessor-aware attacks recover the same efficacy as when attacking the model alone. The code can be found at https://github.com/google-research/preprocessor-aware-black-box-attack.

Daphne Ippolito, Florian Tramer, Milad Nasr, Chiyuan Zhang, Matthew Jagielski, Katherine Lee, Christopher Choquette Choo, Nicholas Carlini. Proceedings of the 16th International Natural Language Generation Conference. 2023.

Large language models are now tuned to align with the goals of their creators, namely to be "helpful and harmless." These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this work, we study adversarial alignment, and ask to what extent these models remain aligned when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). These inputs are designed to cause the model to emit harmful content that would otherwise be prohibited. We show that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, we can find adversarial inputs with brute force. As a result, the failure of current attacks should not be seen as proof that aligned text models remain aligned under adversarial inputs. However the recent trend in large-scale ML models is multimodal models that allow users to provide images that influence the text that is generated. We show these models can be easily attacked, i.e., induced to perform arbitrary un-aligned behavior through adversarial perturbation of the input image. We conjecture that improved NLP attacks may demonstrate this same level of adversarial control over text-only models.

Model distillation is frequently proposed as a technique to reduce the privacy leakage of machine learning. These empirical privacy defenses rely on the intuition that distilled ``student'' models protect the privacy of training data, as they only interact with this data indirectly through a ``teacher'' model. In this work, we design membership inference attacks to systematically study the privacy provided by knowledge distillation to both the teacher and student training sets. Our new attacks show that distillation alone provides only limited privacy across a number of domains. We explain the success of our attacks on distillation by showing that membership inference attacks on a private dataset can succeed even if the target model is never queried on any actual training points, but only on inputs whose predictions are highly influenced by training data. Finally, we show that our attacks are strongest when student and teacher sets are similar, or when the attacker can poison the teacher set.

Modern neural language models that are widely used in various NLP tasks risk memorizing sensitive information from their training data. Understanding this memorization is important in real world applications and also from a learning-theoretical perspective. An open question in previous studies of language model memorization is how to filter out ``common'' memorization. In fact, most memorization criteria strongly correlate with the number of occurrences in the training set, capturing memorized familiar phrases, public knowledge, templated texts, or other repeated data. We formulate a notion of counterfactual memorization which characterizes how a model's predictions change if a particular document is omitted during training. We identify and study counterfactually-memorized training examples in standard text datasets. We estimate the influence of each memorized training example on the validation set and on generated texts, showing how this can provide direct evidence of the source of memorization at test time.

Property inference attacks allow an adversary to extract global properties of the training dataset from a machine learning model. Such attacks have privacy implications for data owners sharing their datasets to train machine learning models. Several existing approaches for property inference attacks against deep neural networks have been proposed [1] –[3], but they all rely on the attacker training a large number of shadow models, which induces a large computational overhead.In this paper, we consider the setting of property inference attacks in which the attacker can poison a subset of the training dataset and query the trained target model. Motivated by our theoretical analysis of model confidences under poisoning, we design an efficient property inference attack, SNAP, which obtains higher attack success and requires lower amounts of poisoning than the state-of-the-art poisoning-based property inference attack by Mahloujifar et al. [3]. For example, on the Census dataset, SNAP achieves 34% higher success rate than [3] while being 56.5× faster. We also extend our attack to infer whether a certain property was present at all during training and estimate the exact proportion of a property of interest efficiently. We evaluate our attack on several properties of varying proportions from four datasets and demonstrate SNAP’s generality and effectiveness.

Recent years have seen a dramatic increase in applications of Artificial Intelligence (AI), Machine Learning (ML), and data mining to security and privacy problems. The analytic tools and intelligent behavior provided by these techniques make AI and ML increasingly important for autonomous real-time analysis and decision making in domains with a wealth of data or that require quick reactions to constantly changing situations. The use of learning methods in security-sensitive domains, in which adversaries may attempt to mislead or evade intelligent machines, creates new frontiers for security research. The recent widespread adoption of "deep learning" techniques, whose security properties are difficult to reason about directly, has only added to the importance of this research. In addition, data mining and machine learning techniques create a wealth of privacy issues, due to the abundance and accessibility of data. The AISec workshop provides a venue for presenting and discussing new developments in the intersection of security and privacy with AI and machine learning.

We introduce a new class of attacks on machine learning models. We show that an adversary who can poison a training dataset can cause models trained on this dataset to leak significant private details of training points belonging to other parties. Our active inference attacks connect two independent lines of work targeting the integrity and privacy of machine learning training data.

Making classifiers robust to adversarial examples is challenging. Thus, many works tackle the seemingly easier task of detecting perturbed inputs. We show a barrier towards this goal. We prove a hardness reduction between detection and classification of adversarial examples: given a robust detector for attacks at distance $\epsilon$ (in some metric), we show how to build a similarly robust (but inefficient) classifier for attacks at distance $\epsilon/2$. Our reduction is computationally inefficient, but preserves the data complexity of the original detector. The reduction thus cannot be directly used to build practical classifiers. Instead, it is a useful sanity check to test whether empirical detection results imply something much stronger than the authors presumably anticipated (namely a highly robust and data-efficient classifier). To illustrate, we revisit $14$ empirical detector defenses published over the past years. For $12/14$ defenses, we show that the claimed detection results imply an inefficient classifier with robustness far beyond the state-of-the-art— thus casting some doubts on the results’ validity. Finally, we show that our reduction applies in both directions: a robust classifier for attacks at distance $\epsilon/2$ implies an inefficient robust detector at distance $\epsilon$. Thus, we argue that robust classification and robust detection should be regarded as (near)-equivalent problems, if we disregard their computational complexity.

Machine learning models trained on private datasets have been shown to leak their private data. Recent work has found that the average data point is rarely leaked---it is often the outlier samples that are subject to memorization and, consequently, leakage. We demonstrate and analyze an Onion Effect of memorization: removing the "layer" of outlier points that are most vulnerable to a privacy attack exposes a new layer of previously-safe points to the same attack. We perform several experiments that are consistent with this hypothesis. For example, we show that for membership inference attacks, when the layer of easiest-to-attack examples is removed, another layer below becomes easy-to-attack. The existence of this effect has various consequences. For example, it suggests that proposals to defend against memorization without training with rigorous privacy guarantees are unlikely to be effective. Further, it suggests that privacy-enhancing technologies such as machine unlearning could actually harm the privacy of other users.

Hundreds of defenses have been proposed to make deep neural networks robust against minimal (adversarial) input perturbations. However, only a handful of these defenses held up their claims because correctly evaluating robustness is extremely challenging: Weak attacks often fail to find adversarial examples even if they unknowingly exist, thereby making a vulnerable network look robust. In this paper, we propose a test to identify weak attacks and, thus, weak defense evaluations. Our test slightly modifies a neural network to guarantee the existence of an adversarial example for every sample. Consequentially, any correct attack must succeed in breaking this modified network. For eleven out of thirteen previously-published defenses, the original evaluation of the defense fails our test, while stronger attacks that break these defenses pass it. We hope that attack unit tests - such as ours - will be a major component in future robustness evaluations and increase confidence in an empirical field that is currently riddled with skepticism.

A membership inference attack allows an adversary to query a trained machine learning model to predict whether or not a particular example was contained in the model’s training dataset. These attacks are currently evaluated using average-case “accuracy” metrics that fail to characterize whether the attack can confidently identify any members of the training set. We argue that attacks should instead be evaluated by computing their true-positive rate at low (e.g., ≤ 0.1%) false-positive rates, and find most prior attacks perform poorly when evaluated in this way. To address this we develop a Likelihood Ratio Attack (LiRA) that carefully combines multiple ideas from the literature. Our attack is $10\times$ more powerful at low false-positive rates, and also strictly dominates prior attacks on existing metrics.

Membership inference is one of the simplest privacy threats faced by machine learning models that are trained on private sensitive data. In this attack, an adversary infers whether a particular point was used to train the model, or not, by observing the model’s predictions. Whereas current attack methods all require access to the model’s predicted confidence score, we introduce a label-only attack that instead evaluates the robustness of the model’s predicted (hard) labels under perturbations of the input, to infer membership. Our label-only attack is not only as-effective as attacks requiring access to confidence scores, it also demonstrates that a class of defenses against membership inference, which we call “confidence masking” because they obfuscate the confidence scores to thwart attacks, are insufficient to prevent the leakage of private information. Our experiments show that training with differential privacy or strong L2 regularization are the only current defenses that meaningfully decrease leakage of private information, even for points that are outliers of the training distribution.

We consider the privacy-preserving machine learning (ML) setting where the trained model must satisfy differential privacy (DP) with respect to the labels of the training examples. We propose two novel approaches based on, respectively, the Laplace mechanism and the PATE framework, and demonstrate their effectiveness on standard benchmarks.While recent work by Ghazi et al. proposed Label DP schemes based on a randomized response mechanism, we argue that additive Laplace noise coupled with Bayesian inference (ALIBI) is a better fit for typical ML tasks. Moreover, we show how to achieve very strong privacy levels in some regimes, with our adaptation of the PATE framework that builds on recent advances in semi-supervised learning.We complement theoretical analysis of our algorithms' privacy guarantees with empirical evaluation of their memorization properties. Our evaluation suggests that comparing different algorithms according to their provable DP guarantees can be misleading and favor a less private algorithm with a tighter analysis.Code for implementation of algorithms and memorization attacks is available from https://github.com/facebookresearch/labeldpantipodes.

A private machine learning algorithm hides as much as possible about its training data while still preserving accuracy. In this work, we study whether a non-private learning algorithm can be made private by relying on an instance-encoding mechanism that modifies the training inputs before feeding them to a normal learner. We formalize both the notion of instance encoding and its privacy by providing two attack models. We first prove impossibility results for achieving a (stronger) model. Next, we demonstrate practical attacks in the second (weaker) attack model on InstaHide, a recent proposal by Huang, Song, Li and Arora [ICML’20] that aims to use instance encoding for privacy.

Adversarial examples are malicious inputs crafted to induce misclassification. Commonly studied \emph{sensitivity-based} adversarial examples introduce semantically-small changes to an input that result in a different model prediction. This paper studies a complementary failure mode, \emph{invariance-based} adversarial examples, that introduce minimal semantic changes that modify an input’s true label yet preserve the model’s prediction. We demonstrate fundamental tradeoffs between these two types of adversarial examples. We show that defenses against sensitivity-based attacks actively harm a model’s accuracy on invariance-based attacks, and that new approaches are needed to resist both attack types. In particular, we break state-of-the-art adversarially-trained and \emph{certifiably-robust} models by generating small perturbations that the models are (provably) robust to, yet that change an input’s class according to human labelers. Finally, we formally show that the existence of excessively invariant classifiers arises from the presence of \emph{overly-robust} predictive features in standard datasets.

Adaptive attacks have (rightfully) become the de facto standard for evaluating defenses to adversarial examples. We find, however, that typical adaptive evaluations are incomplete. We demonstrate that 13 defenses recently published at ICLR, ICML and NeurIPS---and which illustrate a diverse set of defense strategies---can be circumvented despite attempting to perform evaluations using adaptive attacks. While prior evaluation papers focused mainly on the end result---showing that a defense was ineffective---this paper focuses on laying out the methodology and the approach necessary to perform an adaptive attack. Some of our attack strategies are generalizable, but no single strategy would have been sufficient for all defenses. This underlines our key message that adaptive attacks cannot be automated and always require careful and appropriate tuning to a given defense. We hope that these analyses will serve as guidance on how to properly perform adaptive attacks against defenses to adversarial examples, and thus will allow the community to make further progress in building more robust models.

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.

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. We propose new multi-perturbation adversarial training schemes, as well as an efficient attack for the $\ell_1$-norm, and use these to show that models trained against multiple attacks fail to achieve robustness competitive with that of models trained on each attack individually. In particular, we find that adversarial training with first-order $\ell_\infty, \ell_1$ and $\ell_2$ attacks on MNIST achieves merely $50\%$ robust accuracy, partly because of gradient-masking. Finally, we propose affine attacks that linearly interpolate between perturbation types and further degrade the accuracy of adversarially trained models.

In a world where traditional notions of privacy are increasingly challenged by the myriad companies that collect and analyze our data, it is important that decision-making entities are held accountable for unfair treatments arising from irresponsible data usage. Unfortunately, a lack of appropriate methodologies and tools means that even identifying unfair or discriminatory effects can be a challenge in practice. We introduce the unwarranted associations (UA) framework, a principled methodology for the discovery of unfair, discriminatory, or offensive user treatment in data-driven applications. The UA framework unifies and rationalizes a number of prior attempts at formalizing algorithmic fairness. It uniquely combines multiple investigative primitives and fairness metrics with broad applicability, granular exploration of unfair treatment in user subgroups, and incorporation of natural notions of utility that may account for observed disparities. We instantiate the UA framework in FairTest, the first comprehensive tool that helps developers check data-driven applications for unfair user treatment. It enables scalable and statistically rigorous investigation of associations between application outcomes (such as prices or premiums) and sensitive user attributes (such as race or gender). Furthermore, FairTest provides debugging capabilities that let programmers rule out potential confounders for observed unfair effects. We report on use of FairTest to investigate and in some cases address disparate impact, offensive labeling, and uneven rates of algorithmic error in four data-driven applications. As examples, our results reveal subtle biases against older populations in the distribution of error in a predictive health application and offensive racial labeling in an image tagger.

Trusted hardware systems, such as Intel's new SGX instruction set architecture extension, aim to provide strong confidentiality and integrity assurances for applications. Recent work, however, raises serious concerns about the vulnerability of such systems to side-channel attacks. We propose, formalize, and explore a cryptographic primitive called a Sealed-Glass Proof (SGP) that models computation possible in an isolated execution environment with unbounded leakage, and thus in the face of arbitrary side-channels. A SGP specifically models the capabilities of trusted hardware that can attest to correct execution of a piece of code, but whose execution is transparent, meaning that an application's secrets and state are visible to other processes on the same host. Despite this strong threat model, we show that SGPs enable a range of practical applications. Our key observation is that SGPs permit safe verifiable computing in zero-knowledge, as data leakage results only in the prover learning her own secrets. Among other applications, we describe the implementation of an end-to-end bug bounty (or zero-day solicitation) platform that couples a SGX-based SGP with a smart contract. Our platform enables a marketplace that achieves fair exchange, protects against unfair bounty withdrawals, and resists denial-of-service attacks by dishonest sellers. We also consider a slight relaxation of the SGP model that permits black-box modules instantiating minimal, side-channel resistant primitives, yielding a still broader range of applications. Our work shows how trusted hardware systems such as SGX can support trustworthy applications even in the presence of side channels.

The Global Alliance for Genomics and Health (GA4GH) created the Beacon Project as a means of testing the willingness of data holders to share genetic data in the simplest technical context-a query for the presence of a specified nucleotide at a given position within a chromosome. Each participating site (or "beacon") is responsible for assuring that genomic data are exposed through the Beacon service only with the permission of the individual to whom the data pertains and in accordance with the GA4GH policy and standards.While recognizing the inference risks associated with large-scale data aggregation, and the fact that some beacons contain sensitive phenotypic associations that increase privacy risk, the GA4GH adjudged the risk of re-identification based on the binary yes/no allele-presence query responses as acceptable. However, recent work demonstrated that, given a beacon with specific characteristics (including relatively small sample size and an adversary who possesses an individual's whole genome sequence), the individual's membership in a beacon can be inferred through repeated queries for variants present in the individual's genome.In this paper, we propose three practical strategies for reducing re-identification risks in beacons. The first two strategies manipulate the beacon such that the presence of rare alleles is obscured; the third strategy budgets the number of accesses per user for each individual genome. Using a beacon containing data from the 1000 Genomes Project, we demonstrate that the proposed strategies can effectively reduce re-identification risk in beacon-like datasets.

Abstract In the past few years, we have witnessed a rise in the popularity of ride-hailing services (RHSs), an online marketplace that enables accredited drivers to use their own cars to drive ride-hailing users. Unlike other transportation services, RHSs raise significant privacy concerns, as providers are able to track the precise mobility patterns of millions of riders worldwide. We present the first survey and analysis of the privacy threats in RHSs. Our analysis exposes high-risk privacy threats that do not occur in conventional taxi services. Therefore, we propose PrivateRide, a privacy-enhancing and practical solution that offers anonymity and location privacy for riders, and protects drivers’ information from harvesting attacks. PrivateRide lowers the high-risk privacy threats in RHSs to a level that is at least as low as that of many taxi services. Using real data-sets from Uber and taxi rides, we show that PrivateRide significantly enhances riders’ privacy, while preserving tangible accuracy in ride matching and fare calculation, with only negligible effects on convenience. Moreover, by using our Android implementation for experimental evaluations, we show that PrivateRide’s overhead during ride setup is negligible. In short, we enable privacy-conscious riders to achieve levels of privacy that are not possible in current RHSs and even in some conventional taxi services, thereby offering a potential business differentiator.

Differential privacy (DP) has become widely accepted as a rigorous definition of data privacy, with stronger privacy guarantees than traditional statistical methods. However, recent studies have shown that for reasonable privacy budgets, differential privacy significantly affects the expected utility. Many alternative privacy notions which aim at relaxing DP have since been proposed, with the hope of providing a better tradeoff between privacy and utility.