Yves-Alexandre de Montjoye

Alexander Panfilov, Peter Romov, Igor Shilov et al.

LLM agents like Claude Code can not only write code but also be used for autonomous AI research and engineering \citep{rank2026posttrainbench, novikov2025alphaevolve}. We show that an \emph{autoresearch}-style pipeline \citep{karpathy2026autoresearch} powered by Claude Code discovers novel white-box adversarial attack \textit{algorithms} that \textbf{significantly outperform all existing (30+) methods} in jailbreaking and prompt injection evaluations. Starting from existing attack implementations, such as GCG~\citep{zou2023universal}, the agent iterates to produce new algorithms achieving up to 40\% attack success rate on CBRN queries against GPT-OSS-Safeguard-20B, compared to $\leq$10\% for existing algorithms (\Cref{fig:teaser}, left). The discovered algorithms generalize: attacks optimized on surrogate models transfer directly to held-out models, achieving \textbf{100\% ASR against Meta-SecAlign-70B} \citep{chen2025secalign} versus 56\% for the best baseline (\Cref{fig:teaser}, middle). Extending the findings of~\cite{carlini2025autoadvexbench}, our results are an early demonstration that incremental safety and security research can be automated using LLM agents. White-box adversarial red-teaming is particularly well-suited for this: existing methods provide strong starting points, and the optimization objective yields dense, quantitative feedback. We release all discovered attacks alongside baseline implementations and evaluation code at https://github.com/romovpa/claudini.

Matthieu Meeus, Shubham Jain, Marek Rei et al.

With large language models (LLMs) poised to become embedded in our daily lives, questions are starting to be raised about the data they learned from. These questions range from potential bias or misinformation LLMs could retain from their training data to questions of copyright and fair use of human-generated text. However, while these questions emerge, developers of the recent state-of-the-art LLMs become increasingly reluctant to disclose details on their training corpus. We here introduce the task of document-level membership inference for real-world LLMs, i.e. inferring whether the LLM has seen a given document during training or not. First, we propose a procedure for the development and evaluation of document-level membership inference for LLMs by leveraging commonly used data sources for training and the model release date. We then propose a practical, black-box method to predict document-level membership and instantiate it on OpenLLaMA-7B with both books and academic papers. We show our methodology to perform very well, reaching an AUC of 0.856 for books and 0.678 for papers. We then show our approach to outperform the sentence-level membership inference attacks used in the privacy literature for the document-level membership task. We further evaluate whether smaller models might be less sensitive to document-level inference and show OpenLLaMA-3B to be approximately as sensitive as OpenLLaMA-7B to our approach. Finally, we consider two mitigation strategies and find the AUC to slowly decrease when only partial documents are considered but to remain fairly high when the model precision is reduced. Taken together, our results show that accurate document-level membership can be inferred for LLMs, increasing the transparency of technology poised to change our lives.

Ana-Maria Cretu, Florimond Houssiau, Antoine Cully et al.

Although query-based systems (QBS) have become one of the main solutions to share data anonymously, building QBSes that robustly protect the privacy of individuals contributing to the dataset is a hard problem. Theoretical solutions relying on differential privacy guarantees are difficult to implement correctly with reasonable accuracy, while ad-hoc solutions might contain unknown vulnerabilities. Evaluating the privacy provided by QBSes must thus be done by evaluating the accuracy of a wide range of privacy attacks. However, existing attacks require time and expertise to develop, need to be manually tailored to the specific systems attacked, and are limited in scope. In this paper, we develop QuerySnout (QS), the first method to automatically discover vulnerabilities in QBSes. QS takes as input a target record and the QBS as a black box, analyzes its behavior on one or more datasets, and outputs a multiset of queries together with a rule to combine answers to them in order to reveal the sensitive attribute of the target record. QS uses evolutionary search techniques based on a novel mutation operator to find a multiset of queries susceptible to lead to an attack, and a machine learning classifier to infer the sensitive attribute from answers to the queries selected. We showcase the versatility of QS by applying it to two attack scenarios, three real-world datasets, and a variety of protection mechanisms. We show the attacks found by QS to consistently equate or outperform, sometimes by a large margin, the best attacks from the literature. We finally show how QS can be extended to QBSes that require a budget, and apply QS to a simple QBS based on the Laplace mechanism. Taken together, our results show how powerful and accurate attacks against QBSes can already be found by an automated system, allowing for highly complex QBSes to be automatically tested "at the pressing of a button".

Matthieu Meeus, Florent Guépin, Ana-Maria Cretu et al.

Synthetic data is seen as the most promising solution to share individual-level data while preserving privacy. Shadow modeling-based Membership Inference Attacks (MIAs) have become the standard approach to evaluate the privacy risk of synthetic data. While very effective, they require a large number of datasets to be created and models trained to evaluate the risk posed by a single record. The privacy risk of a dataset is thus currently evaluated by running MIAs on a handful of records selected using ad-hoc methods. We here propose what is, to the best of our knowledge, the first principled vulnerable record identification technique for synthetic data publishing, leveraging the distance to a record's closest neighbors. We show our method to strongly outperform previous ad-hoc methods across datasets and generators. We also show evidence of our method to be robust to the choice of MIA and to specific choice of parameters. Finally, we show it to accurately identify vulnerable records when synthetic data generators are made differentially private. The choice of vulnerable records is as important as more accurate MIAs when evaluating the privacy of synthetic data releases, including from a legal perspective. We here propose a simple yet highly effective method to do so. We hope our method will enable practitioners to better estimate the risk posed by synthetic data publishing and researchers to fairly compare ever improving MIAs on synthetic data.

Florent Guépin, Matthieu Meeus, Ana-Maria Cretu et al.

Synthetic data is emerging as one of the most promising solutions to share individual-level data while safeguarding privacy. While membership inference attacks (MIAs), based on shadow modeling, have become the standard to evaluate the privacy of synthetic data, they currently assume the attacker to have access to an auxiliary dataset sampled from a similar distribution as the training dataset. This is often seen as a very strong assumption in practice, especially as the proposed main use cases for synthetic tabular data (e.g. medical data, financial transactions) are very specific and don't have any reference datasets directly available. We here show how this assumption can be removed, allowing for MIAs to be performed using only the synthetic data. Specifically, we developed three different scenarios: (S1) Black-box access to the generator, (S2) only access to the released synthetic dataset and (S3) a theoretical setup as upper bound for the attack performance using only synthetic data. Our results show that MIAs are still successful, across two real-world datasets and two synthetic data generators. These results show how the strong hypothesis made when auditing synthetic data releases - access to an auxiliary dataset - can be relaxed, making the attacks more realistic in practice.

Ana-Maria Cretu, Daniel Jones, Yves-Alexandre de Montjoye et al.

Machine learning models have been shown to leak sensitive information about their training datasets. Models are increasingly deployed on devices, raising concerns that white-box access to the model parameters increases the attack surface compared to black-box access which only provides query access. Directly extending the shadow modelling technique from the black-box to the white-box setting has been shown, in general, not to perform better than black-box only attacks. A potential reason is misalignment, a known characteristic of deep neural networks. In the shadow modelling context, misalignment means that, while the shadow models learn similar features in each layer, the features are located in different positions. We here present the first systematic analysis of the causes of misalignment in shadow models and show the use of a different weight initialisation to be the main cause. We then extend several re-alignment techniques, previously developed in the model fusion literature, to the shadow modelling context, where the goal is to re-align the layers of a shadow model to those of the target model. We show re-alignment techniques to significantly reduce the measured misalignment between the target and shadow models. Finally, we perform a comprehensive evaluation of white-box membership inference attacks (MIA). Our analysis reveals that internal layer activation-based MIAs suffer strongly from shadow model misalignment, while gradient-based MIAs are only sometimes significantly affected. We show that re-aligning the shadow models strongly improves the former's performance and can also improve the latter's performance, although less frequently. Taken together, our results highlight that on-device deployment increases the attack surface and that the newly available information can be used to build more powerful attacks.

Bozhidar Stevanoski, Ana-Maria Cretu, Yves-Alexandre de Montjoye

Query-based systems (QBSs) are one of the key approaches for sharing data. QBSs allow analysts to request aggregate information from a private protected dataset. Attacks are a crucial part of ensuring QBSs are truly privacy-preserving. The development and testing of attacks is however very labor-intensive and unable to cope with the increasing complexity of systems. Automated approaches have been shown to be promising but are currently extremely computationally intensive, limiting their applicability in practice. We here propose QueryCheetah, a fast and effective method for automated discovery of privacy attacks against QBSs. We instantiate QueryCheetah on attribute inference attacks and show it to discover stronger attacks than previous methods while being 18 times faster than the state-of-the-art automated approach. We then show how QueryCheetah allows system developers to thoroughly evaluate the privacy risk, including for various attacker strengths and target individuals. We finally show how QueryCheetah can be used out-of-the-box to find attacks in larger syntaxes and workarounds around ad-hoc defenses.

Florimond Houssiau, Vincent Schellekens, Antoine Chatalic et al.

Differential privacy is the standard privacy definition for performing analyses over sensitive data. Yet, its privacy budget bounds the number of tasks an analyst can perform with reasonable accuracy, which makes it challenging to deploy in practice. This can be alleviated by private sketching, where the dataset is compressed into a single noisy sketch vector which can be shared with the analysts and used to perform arbitrarily many analyses. However, the algorithms to perform specific tasks from sketches must be developed on a case-by-case basis, which is a major impediment to their use. In this paper, we introduce the generic moment-to-moment (M$^2$M) method to perform a wide range of data exploration tasks from a single private sketch. Among other things, this method can be used to estimate empirical moments of attributes, the covariance matrix, counting queries (including histograms), and regression models. Our method treats the sketching mechanism as a black-box operation, and can thus be applied to a wide variety of sketches from the literature, widening their ranges of applications without further engineering or privacy loss, and removing some of the technical barriers to the wider adoption of sketches for data exploration under differential privacy. We validate our method with data exploration tasks on artificial and real-world data, and show that it can be used to reliably estimate statistics and train classification models from private sketches.

Nataša Krčo, Zexi Yao, Matthieu Meeus et al.

Data containing personal information is increasingly used to train, fine-tune, or query Large Language Models (LLMs). Text is typically scrubbed of identifying information prior to use, often with tools such as Microsoft's Presidio or Anthropic's PII purifier. These tools have traditionally been evaluated on their ability to remove specific identifiers (e.g., names), yet their effectiveness at preventing re-identification remains unclear. We introduce RAT-Bench, a comprehensive benchmark for text anonymization tools based on re-identification risk. Using U.S. demographic statistics, we generate synthetic text containing various direct and indirect identifiers across domains, languages, and difficulty levels. We evaluate a range of NER- and LLM-based text anonymization tools and, based on the attributes an LLM-based attacker is able to correctly infer from the anonymized text, we report the risk of re-identification in the U.S. population, while properly accounting for the disparate impact of identifiers. We find that, while capabilities vary widely, even the best tools are far from perfect in particular when direct identifiers are not written in standard ways and when indirect identifiers enable re-identification. Overall we find LLM-based anonymizers, including new iterative anonymizers, to provide a better privacy-utility trade-off albeit at a higher computational cost. Importantly, we also find them to work well across languages. We conclude with recommendations for future anonymization tools and will release the benchmark and encourage community efforts to expand it, in particular to other geographies.

Matthieu Meeus, Igor Shilov, Manuel Faysse et al. · meta-ai

Questions of fair use of copyright-protected content to train Large Language Models (LLMs) are being actively debated. Document-level inference has been proposed as a new task: inferring from black-box access to the trained model whether a piece of content has been seen during training. SOTA methods however rely on naturally occurring memorization of (part of) the content. While very effective against models that memorize significantly, we hypothesize--and later confirm--that they will not work against models that do not naturally memorize, e.g. medium-size 1B models. We here propose to use copyright traps, the inclusion of fictitious entries in original content, to detect the use of copyrighted materials in LLMs with a focus on models where memorization does not naturally occur. We carefully design a randomized controlled experimental setup, inserting traps into original content (books) and train a 1.3B LLM from scratch. We first validate that the use of content in our target model would be undetectable using existing methods. We then show, contrary to intuition, that even medium-length trap sentences repeated a significant number of times (100) are not detectable using existing methods. However, we show that longer sequences repeated a large number of times can be reliably detected (AUC=0.75) and used as copyright traps. Beyond copyright applications, our findings contribute to the study of LLM memorization: the randomized controlled setup enables us to draw causal relationships between memorization and certain sequence properties such as repetition in model training data and perplexity.

Jamie Hayes, Ilia Shumailov, Christopher A. Choquette-Choo et al. · deepmind

State-of-the-art membership inference attacks (MIAs) typically require training many reference models, making it difficult to scale these attacks to large pre-trained language models (LLMs). As a result, prior research has either relied on weaker attacks that avoid training references (e.g., fine-tuning attacks), or on stronger attacks applied to small models and datasets. However, weaker attacks have been shown to be brittle and insights from strong attacks in simplified settings do not translate to today's LLMs. These challenges prompt an important question: are the limitations observed in prior work due to attack design choices, or are MIAs fundamentally ineffective on LLMs? We address this question by scaling LiRA--one of the strongest MIAs--to GPT-2 architectures ranging from 10M to 1B parameters, training references on over 20B tokens from the C4 dataset. Our results advance the understanding of MIAs on LLMs in four key ways. While (1) strong MIAs can succeed on pre-trained LLMs, (2) their effectiveness, remains limited (e.g., AUC<0.7) in practical settings. (3) Even when strong MIAs achieve better-than-random AUC, aggregate metrics can conceal substantial per-sample MIA decision instability: due to training randomness, many decisions are so unstable that they are statistically indistinguishable from a coin flip. Finally, (4) the relationship between MIA success and related LLM privacy metrics is not as straightforward as prior work has suggested.

Zexi Yao, Nataša Krčo, Georgi Ganev et al.

Synthetic data has become an increasingly popular way to share data without revealing sensitive information. Though Membership Inference Attacks (MIAs) are widely considered the gold standard for empirically assessing the privacy of a synthetic dataset, practitioners and researchers often rely on simpler proxy metrics such as Distance to Closest Record (DCR). These metrics estimate privacy by measuring the similarity between the training data and generated synthetic data. This similarity is also compared against that between the training data and a disjoint holdout set of real records to construct a binary privacy test. If the synthetic data is not more similar to the training data than the holdout set is, it passes the test and is considered private. In this work we show that, while computationally inexpensive, DCR and other distance-based metrics fail to identify privacy leakage. Across multiple datasets and both classical models such as Baynet and CTGAN and more recent diffusion models, we show that datasets deemed private by proxy metrics are highly vulnerable to MIAs. We similarly find both the binary privacy test and the continuous measure based on these metrics to be uninformative of actual membership inference risk. We further show that these failures are consistent across different metric hyperparameter settings and record selection methods. Finally, we argue DCR and other distance-based metrics to be flawed by design and show a example of a simple leakage they miss in practice. With this work, we hope to motivate practitioners to move away from proxy metrics to MIAs as the rigorous, comprehensive standard of evaluating privacy of synthetic data, in particular to make claims of datasets being legally anonymous.

Joseph Pollock, Igor Shilov, Euodia Dodd et al.

Membership inference attacks (MIAs) are widely used to empirically assess privacy risks in machine learning models, both providing model-level vulnerability metrics and identifying the most vulnerable training samples. State-of-the-art methods, however, require training hundreds of shadow models with the same architecture as the target model. This makes the computational cost of assessing the privacy of models prohibitive for many practical applications, particularly when used iteratively as part of the model development process and for large models. We propose a novel approach for identifying the training samples most vulnerable to membership inference attacks by analyzing artifacts naturally available during the training process. Our method, Loss Trace Interquartile Range (LT-IQR), analyzes per-sample loss trajectories collected during model training to identify high-risk samples without requiring any additional model training. Through experiments on standard benchmarks, we demonstrate that LT-IQR achieves 92% precision@k=1% in identifying the samples most vulnerable to state-of-the-art MIAs. This result holds across datasets and model architectures with LT-IQR outperforming both traditional vulnerability metrics, such as loss, and lightweight MIAs using few shadow models. We also show LT-IQR to accurately identify points vulnerable to multiple MIA methods and perform ablation studies. We believe LT-IQR enables model developers to identify vulnerable training samples, for free, as part of the model development process. Our results emphasize the potential of artifact-based methods to efficiently evaluate privacy risks.

Pascal Epple, Igor Shilov, Bozhidar Stevanoski et al.

Generative Artificial Intelligence (Gen-AI) models are increasingly used to produce content across domains, including text, images, and audio. While these models represent a major technical breakthrough, they gain their generative capabilities from being trained on enormous amounts of human-generated content, which often includes copyrighted material. In this work, we investigate whether audio watermarking techniques can be used to detect an unauthorized usage of content to train a music generation model. We compare outputs generated by a model trained on watermarked data to a model trained on non-watermarked data. We study factors that impact the model's generation behaviour: the watermarking technique, the proportion of watermarked samples in the training set, and the robustness of the watermarking technique against the model's tokenizer. Our results show that audio watermarking techniques, including some that are imperceptible to humans, can lead to noticeable shifts in the model's outputs. We also study the robustness of a state-of-the-art watermarking technique to removal techniques.

Igor Shilov, Matthieu Meeus, Yves-Alexandre de Montjoye

As Large Language Models (LLMs) become widely adopted, understanding how they learn from, and memorize, training data becomes crucial. Memorization in LLMs is widely assumed to only occur as a result of sequences being repeated in the training data. Instead, we show that LLMs memorize by assembling information from similar sequences, a phenomena we call mosaic memory. We show major LLMs to exhibit mosaic memory, with fuzzy duplicates contributing to memorization as much as 0.8 of an exact duplicate and even heavily modified sequences contributing substantially to memorization. Despite models display reasoning capabilities, we somewhat surprisingly show memorization to be predominantly syntactic rather than semantic. We finally show fuzzy duplicates to be ubiquitous in real-world data, untouched by deduplication techniques. Taken together, our results challenge widely held beliefs and show memorization to be a more complex, mosaic process, with real-world implications for privacy, confidentiality, model utility and evaluation.

Ana-Maria Cretu, Miruna Rusu, Yves-Alexandre de Montjoye

Smart meters, devices measuring the electricity and gas consumption of a household, are currently being deployed at a fast rate throughout the world. The data they collect are extremely useful, including in the fight against climate change. However, these data and the information that can be inferred from them are highly sensitive. Re-pseudonymization, i.e., the frequent replacement of random identifiers over time, is widely used to share smart meter data while mitigating the risk of re-identification. We here show how, in spite of re-pseudonymization, households' consumption records can be pieced together with high accuracy in large-scale datasets. We propose the first deep learning-based profiling attack against re-pseudonymized smart meter data. Our attack combines neural network embeddings, which are used to extract features from weekly consumption records and are tailored to the smart meter identification task, with a nearest neighbor classifier. We evaluate six neural networks architectures as the embedding model. Our results suggest that the Transformer and CNN-LSTM architectures vastly outperform previous methods as well as other architectures, successfully identifying the correct household 73.4% of the time among 5139 households based on electricity and gas consumption records (54.5% for electricity only). We further show that the features extracted by the embedding model maintain their effectiveness when transferred to a set of users disjoint from the one used to train the model. Finally, we extensively evaluate the robustness of our results. Taken together, our results strongly suggest that even frequent re-pseudonymization strategies can be reversed, strongly limiting their ability to prevent re-identification in practice.

Xiaoxue Yang, Bozhidar Stevanoski, Matthieu Meeus et al.

Large language models (LLMs) are increasingly deployed in real-world applications ranging from chatbots to agentic systems, where they are expected to process untrusted data and follow trusted instructions. Failure to distinguish between the two poses significant security risks, exploited by prompt injection attacks, which inject malicious instructions into the data to control model outputs. Model-level defenses have been proposed to mitigate prompt injection attacks. These defenses fine-tune LLMs to ignore injected instructions in untrusted data. We introduce Checkpoint-GCG, a white-box attack against fine-tuning-based defenses. Checkpoint-GCG enhances the Greedy Coordinate Gradient (GCG) attack by leveraging intermediate model checkpoints produced during fine-tuning to initialize GCG, with each checkpoint acting as a stepping stone for the next one to continuously improve attacks. First, we instantiate Checkpoint-GCG to evaluate the robustness of the state-of-the-art defenses in an auditing setup, assuming both (a) full knowledge of the model input and (b) access to intermediate model checkpoints. We show Checkpoint-GCG to achieve up to $96\%$ attack success rate (ASR) against the strongest defense. Second, we relax the first assumption by searching for a universal suffix that would work on unseen inputs, and obtain up to $89.9\%$ ASR against the strongest defense. Finally, we relax both assumptions by searching for a universal suffix that would transfer to similar black-box models and defenses, achieving an ASR of $63.9\%$ against a newly released defended model from Meta.

Euodia Dodd, Nataša Krčo, Igor Shilov et al.

Membership inference attacks (MIAs) have emerged as the standard tool for evaluating the privacy risks of AI models. However, state-of-the-art attacks require training numerous, often computationally expensive, reference models, limiting their practicality. We present a novel approach for estimating model-level vulnerability, the TPR at low FPR, to membership inference attacks without requiring reference models. Empirical analysis shows loss distributions to be asymmetric and heavy-tailed and suggests that most points at risk from MIAs have moved from the tail (high-loss region) to the head (low-loss region) of the distribution after training. We leverage this insight to propose a method to estimate model-level vulnerability from the training and testing distribution alone: using the absence of outliers from the high-loss region as a predictor of the risk. We evaluate our method, the TNR of a simple loss attack, across a wide range of architectures and datasets and show it to accurately estimate model-level vulnerability to the SOTA MIA attack (LiRA). We also show our method to outperform both low-cost (few reference models) attacks such as RMIA and other measures of distribution difference. We finally evaluate the use of non-linear functions to evaluate risk and show the approach to be promising to evaluate the risk in large-language models.

Matthieu Meeus, Igor Shilov, Georgios Kaissis et al.

Machine learning models are known to memorize samples from their training data, raising concerns around privacy and generalization. Counterfactual self-influence is a popular metric to study memorization, quantifying how the model's prediction for a sample changes depending on the sample's inclusion in the training dataset. However, recent work has shown memorization to be affected by factors beyond self-influence, with other training samples, in particular (near-)duplicates, having a large impact. We here study memorization treating counterfactual influence as a distributional quantity, taking into account how all training samples influence how a sample is memorized. For a small language model, we compute the full influence distribution of training samples on each other and analyze its properties. We find that solely looking at self-influence can severely underestimate tangible risks associated with memorization: the presence of (near-)duplicates seriously reduces self-influence, while we find these samples to be (near-)extractable. We observe similar patterns for image classification, where simply looking at the influence distributions reveals the presence of near-duplicates in CIFAR-10. Our findings highlight that memorization stems from complex interactions across training data and is better captured by the full influence distribution than by self-influence alone.

Yifeng Mao, Bozhidar Stevanoski, Yves-Alexandre de Montjoye

Empirical inference attacks are a popular approach for evaluating the privacy risk of data release mechanisms in practice. While an active attack literature exists to evaluate machine learning models or synthetic data release, we currently lack comparable methods for fixed aggregate statistics, in particular when only a limited number of statistics are released. We here propose an inference attack framework against fixed aggregate statistics and an attribute inference attack called DeSIA. We instantiate DeSIA against the U.S. Census PPMF dataset and show it to strongly outperform reconstruction-based attacks. In particular, we show DeSIA to be highly effective at identifying vulnerable users, achieving a true positive rate of 0.14 at a false positive rate of $10^{-3}$. We then show DeSIA to perform well against users whose attributes cannot be verified and when varying the number of aggregate statistics and level of noise addition. We also perform an extensive ablation study of DeSIA and show how DeSIA can be successfully adapted to the membership inference task. Overall, our results show that aggregation alone is not sufficient to protect privacy, even when a relatively small number of aggregates are being released, and emphasize the need for formal privacy mechanisms and testing before aggregate statistics are released.

Vincent Guan, Florent Guépin, Ana-Maria Cretu et al.

Location data is frequently collected from populations and shared in aggregate form to guide policy and decision making. However, the prevalence of aggregated data also raises the privacy concern of membership inference attacks (MIAs). MIAs infer whether an individual's data contributed to the aggregate release. Although effective MIAs have been developed for aggregate location data, these require access to an extensive auxiliary dataset of individual traces over the same locations, which are collected from a similar population. This assumption is often impractical given common privacy practices surrounding location data. To measure the risk of an MIA performed by a realistic adversary, we develop the first Zero Auxiliary Knowledge (ZK) MIA on aggregate location data, which eliminates the need for an auxiliary dataset of real individual traces. Instead, we develop a novel synthetic approach, such that suitable synthetic traces are generated from the released aggregate. We also develop methods to correct for bias and noise, to show that our synthetic-based attack is still applicable when privacy mechanisms are applied prior to release. Using two large-scale location datasets, we demonstrate that our ZK MIA matches the state-of-the-art Knock-Knock (KK) MIA across a wide range of settings, including popular implementations of differential privacy (DP) and suppression of small counts. Furthermore, we show that ZK MIA remains highly effective even when the adversary only knows a small fraction (10%) of their target's location history. This demonstrates that effective MIAs can be performed by realistic adversaries, highlighting the need for strong DP protection.

Matthieu Meeus, Igor Shilov, Shubham Jain et al.

Whether LLMs memorize their training data and what this means, from measuring privacy leakage to detecting copyright violations, has become a rapidly growing area of research. In the last few months, more than 10 new methods have been proposed to perform Membership Inference Attacks (MIAs) against LLMs. Contrary to traditional MIAs which rely on fixed-but randomized-records or models, these methods are mostly trained and tested on datasets collected post-hoc. Sets of members and non-members, used to evaluate the MIA, are constructed using informed guesses after the release of a model. This lack of randomization raises concerns of a distribution shift between members and non-members. In this work, we first extensively review the literature on MIAs against LLMs and show that, while most work focuses on sequence-level MIAs evaluated in post-hoc setups, a range of target models, motivations and units of interest are considered. We then quantify distribution shifts present in 6 datasets used in the literature using a model-less bag of word classifier and show that all datasets constructed post-hoc suffer from strong distribution shifts. These shifts invalidate the claims of LLMs memorizing strongly in real-world scenarios and, potentially, also the methodological contributions of the recent papers based on these datasets. Yet, all hope might not be lost. We introduce important considerations to properly evaluate MIAs against LLMs and discuss, in turn, potential ways forwards: randomized test splits, injections of randomized (unique) sequences, randomized fine-tuning, and several post-hoc control methods. While each option comes with its advantages and limitations, we believe they collectively provide solid grounds to guide MIA development and study LLM memorization. We conclude with an overview of recommended approaches to benchmark sequence-level and document-level MIAs against LLMs.

Matthew Wicker, Philip Sosnin, Igor Shilov et al.

We study private prediction where differential privacy is achieved by adding noise to the outputs of a non-private model. Existing methods rely on noise proportional to the global sensitivity of the model, often resulting in sub-optimal privacy-utility trade-offs compared to private training. We introduce a novel approach for computing dataset-specific upper bounds on prediction sensitivity by leveraging convex relaxation and bound propagation techniques. By combining these bounds with the smooth sensitivity mechanism, we significantly improve the privacy analysis of private prediction compared to global sensitivity-based approaches. Experimental results across real-world datasets in medical image classification and natural language processing demonstrate that our sensitivity bounds are can be orders of magnitude tighter than global sensitivity. Our approach provides a strong basis for the development of novel privacy preserving technologies.

Nataša Krčo, Florent Guépin, Matthieu Meeus et al.

Synthetic data generators and machine learning models can memorize their training data, posing privacy concerns. Membership inference attacks (MIAs) are a standard method of estimating the privacy risk of these systems. The risk of individual records is typically computed by evaluating MIAs in a record-specific privacy game. We analyze the record-specific privacy game commonly used for evaluating attackers under realistic assumptions (the \textit{traditional} game) -- particularly for synthetic tabular data -- and show that it averages a record's privacy risk across datasets. We show this implicitly assumes the dataset a record is part of has no impact on the record's risk, providing a misleading risk estimate when a specific model or synthetic dataset is released. Instead, we propose a novel use of the leave-one-out game, used in existing work exclusively to audit differential privacy guarantees, and call this the \textit{model-seeded} game. We formalize it and show that it provides an accurate estimate of the privacy risk posed by a given adversary for a record in its specific dataset. We instantiate and evaluate the state-of-the-art MIA for synthetic data generators in the traditional and model-seeded privacy games, and show across multiple datasets and models that the two privacy games indeed result in different risk scores, with up to 94\% of high-risk records being overlooked by the traditional game. We further show that records in smaller datasets and models not protected by strong differential privacy guarantees tend to have a larger gap between risk estimates. Taken together, our results show that the model-seeded setup yields a risk estimate specific to a certain model or synthetic dataset released and in line with the standard notion of privacy leakage from prior work, meaningfully different from the dataset-averaged risk provided by the traditional privacy game.

Ana-Maria Creţu, Florent Guépin, Yves-Alexandre de Montjoye

Despite machine learning models being widely used today, the relationship between a model and its training dataset is not well understood. We explore correlation inference attacks, whether and when a model leaks information about the correlations between the input variables of its training dataset. We first propose a model-less attack, where an adversary exploits the spherical parametrization of correlation matrices alone to make an informed guess. Second, we propose a model-based attack, where an adversary exploits black-box model access to infer the correlations using minimal and realistic assumptions. Third, we evaluate our attacks against logistic regression and multilayer perceptron models on three tabular datasets and show the models to leak correlations. We finally show how extracted correlations can be used as building blocks for attribute inference attacks and enable weaker adversaries. Our results raise fundamental questions on what a model does and should remember from its training set.

Shubham Jain, Ana-Maria Cretu, Yves-Alexandre de Montjoye

End-to-end encryption (E2EE) by messaging platforms enable people to securely and privately communicate with one another. Its widespread adoption however raised concerns that illegal content might now be shared undetected. Following the global pushback against key escrow systems, client-side scanning based on perceptual hashing has been recently proposed by tech companies, governments and researchers to detect illegal content in E2EE communications. We here propose the first framework to evaluate the robustness of perceptual hashing-based client-side scanning to detection avoidance attacks and show current systems to not be robust. More specifically, we propose three adversarial attacks--a general black-box attack and two white-box attacks for discrete cosine transform-based algorithms--against perceptual hashing algorithms. In a large-scale evaluation, we show perceptual hashing-based client-side scanning mechanisms to be highly vulnerable to detection avoidance attacks in a black-box setting, with more than 99.9% of images successfully attacked while preserving the content of the image. We furthermore show our attack to generate diverse perturbations, strongly suggesting that straightforward mitigation strategies would be ineffective. Finally, we show that the larger thresholds necessary to make the attack harder would probably require more than one billion images to be flagged and decrypted daily, raising strong privacy concerns. Taken together, our results shed serious doubts on the robustness of perceptual hashing-based client-side scanning mechanisms currently proposed by governments, organizations, and researchers around the world.

Andrea Gadotti, Florimond Houssiau, Luc Rocher et al.

Anonymized data is highly valuable to both businesses and researchers. A large body of research has however shown the strong limits of the de-identification release-and-forget model, where data is anonymized and shared. This has led to the development of privacy-preserving query-based systems. Based on the idea of "sticky noise", Diffix has been recently proposed as a novel query-based mechanism satisfying alone the EU Article~29 Working Party's definition of anonymization. According to its authors, Diffix adds less noise to answers than solutions based on differential privacy while allowing for an unlimited number of queries. This paper presents a new class of noise-exploitation attacks, exploiting the noise added by the system to infer private information about individuals in the dataset. Our first differential attack uses samples extracted from Diffix in a likelihood ratio test to discriminate between two probability distributions. We show that using this attack against a synthetic best-case dataset allows us to infer private information with 89.4% accuracy using only 5 attributes. Our second cloning attack uses dummy conditions that conditionally strongly affect the output of the query depending on the value of the private attribute. Using this attack on four real-world datasets, we show that we can infer private attributes of at least 93% of the users in the dataset with accuracy between 93.3% and 97.1%, issuing a median of 304 queries per user. We show how to optimize this attack, targeting 55.4% of the users and achieving 91.7% accuracy, using a maximum of only 32 queries per user. Our attacks demonstrate that adding data-dependent noise, as done by Diffix, is not sufficient to prevent inference of private attributes. We furthermore argue that Diffix alone fails to satisfy Art. 29 WP's definition of anonymization. [...]

Florimond Houssiau, Piotr Sapiezynski, Laura Radaelli et al.

Despite proportionality being one of the tenets of data protection laws, we currently lack a robust analytical framework to evaluate the reach of modern data collections and the network effects at play. We here propose a graph-theoretic model and notions of node- and edge-observability to quantify the reach of networked data collections. We first prove closed-form expressions for our metrics and quantify the impact of the graph's structure on observability. Second, using our model, we quantify how (1) from 270,000 compromised accounts, Cambridge Analytica collected 68.0M Facebook profiles; (2) from surveilling 0.01\% the nodes in a mobile phone network, a law-enforcement agency could observe 18.6\% of all communications; and (3) an app installed on 1\% of smartphones could monitor the location of half of the London population through close proximity tracing. Better quantifying the reach of data collection mechanisms is essential to evaluate their proportionality.

Giuseppe D'Acquisto, Josep Domingo-Ferrer, Panayiotis Kikiras et al.

The extensive collection and processing of personal information in big data analytics has given rise to serious privacy concerns, related to wide scale electronic surveillance, profiling, and disclosure of private data. To reap the benefits of analytics without invading the individuals' private sphere, it is essential to draw the limits of big data processing and integrate data protection safeguards in the analytics value chain. ENISA, with the current report, supports this approach and the position that the challenges of technology (for big data) should be addressed by the opportunities of technology (for privacy). We first explain the need to shift from "big data versus privacy" to "big data with privacy". In this respect, the concept of privacy by design is key to identify the privacy requirements early in the big data analytics value chain and in subsequently implementing the necessary technical and organizational measures. After an analysis of the proposed privacy by design strategies in the different phases of the big data value chain, we review privacy enhancing technologies of special interest for the current and future big data landscape. In particular, we discuss anonymization, the "traditional" analytics technique, the emerging area of encrypted search and privacy preserving computations, granular access control mechanisms, policy enforcement and accountability, as well as data provenance issues. Moreover, new transparency and access tools in big data are explored, together with techniques for user empowerment and control. Achieving "big data with privacy" is no easy task and a lot of research and implementation is still needed. Yet, it remains a possible task, as long as all the involved stakeholders take the necessary steps to integrate privacy and data protection safeguards in the heart of big data, by design and by default.

Bjarke Felbo, Pål Sundsøy, Alex 'Sandy' Pentland et al.

Mobile phone metadata is increasingly used for humanitarian purposes in developing countries as traditional data is scarce. Basic demographic information is however often absent from mobile phone datasets, limiting the operational impact of the datasets. For these reasons, there has been a growing interest in predicting demographic information from mobile phone metadata. Previous work focused on creating increasingly advanced features to be modeled with standard machine learning algorithms. We here instead model the raw mobile phone metadata directly using deep learning, exploiting the temporal nature of the patterns in the data. From high-level assumptions we design a data representation and convolutional network architecture for modeling patterns within a week. We then examine three strategies for aggregating patterns across weeks and show that our method reaches state-of-the-art accuracy on both age and gender prediction using only the temporal modality in mobile metadata. We finally validate our method on low activity users and evaluate the modeling assumptions.