arXiv subjects
Publications and source records attributed to Josep Domingo-Ferrer.
We present key challenges and future research directions in the security and privacy of agentic AI, based on a horizon-scanning exercise that brought together thirty leading international experts from academia, industry, and government to engage in focused discussions and collaborative exercises on the emerging risks associated with the growing agency of AI.
Privacy protection laws, such as the GDPR, grant individuals the right to request the forgetting of their personal data not only from databases but also from machine learning (ML) models trained on them. Machine unlearning has emerged as a practical means to facilitate model forgetting of data instances seen during training. Although some existing machine unlearning methods guarantee exact forgetting, they are typically costly in computational terms. On the other hand, more affordable methods do not offer forgetting guarantees and are applicable only to specific ML models. In this paper, we present \emph{efficient unlearning with privacy guarantees} (EUPG), a novel machine unlearning framework that offers formal privacy guarantees to individuals whose data are being unlearned. EUPG involves pre-training ML models on data protected using privacy models, and it enables {\em efficient unlearning with the privacy guarantees offered by the privacy models in use}. Through empirical evaluation on four heterogeneous data sets protected with $k$-anonymity and $ε$-differential privacy as privacy models, our approach demonstrates utility and forgetting effectiveness comparable to those of exact unlearning methods, while significantly reducing computational and storage costs. Our code is available at https://github.com/najeebjebreel/EUPG.
Membership inference attacks (MIAs) aim to determine whether a data sample was included in a machine learning (ML) model's training set and have become the de facto standard for measuring privacy leakages in ML. We propose an evaluation framework that defines the conditions under which MIAs constitute a genuine privacy threat, and review representative MIAs against it. We find that, under the realistic conditions defined in our framework, MIAs represent weak privacy threats. Thus, relying on them as a privacy metric in ML can lead to an overestimation of risk and to unnecessary sacrifices in model utility as a consequence of employing too strong defenses.
Membership inference attacks (MIAs) have become the standard tool for evaluating privacy leakage in machine learning (ML). Among them, the Likelihood-Ratio Attack (LiRA) is widely regarded as the state of the art when sufficient shadow models are available. However, prior evaluations have often overstated the effectiveness of LiRA by attacking models overconfident on their training samples, calibrating thresholds on target data, assuming balanced membership priors, and/or overlooking attack reproducibility. We re-evaluate LiRA under a realistic protocol that (i) trains models using anti-overfitting (AOF) and transfer learning (TL), when applicable, to reduce overconfidence as in production models; (ii) calibrates decision thresholds using shadow models and data rather than target data; (iii) measures positive predictive value (PPV, or precision) under shadow-based thresholds and skewed membership priors (pi <= 10%); and (iv) quantifies per-sample membership reproducibility across different seeds and training variations. We find that AOF significantly weakens LiRA, while TL further reduces attack effectiveness while improving model accuracy. Under shadow-based thresholds and skewed priors, LiRA's PPV often drops substantially, especially under AOF or AOF+TL. We also find that thresholded vulnerable sets at extremely low FPR show poor reproducibility across runs, while likelihood-ratio rankings are more stable. These results suggest that LiRA, and likely weaker MIAs, are less effective than previously suggested under realistic conditions, and that reliable privacy auditing requires evaluation protocols that reflect practical training practices, feasible attacker assumptions, and reproducibility considerations. Code is available at https://github.com/najeebjebreel/lira_analysis.
Membership inference attacks (MIAs) against machine learning (ML) models aim to determine whether a given data point was part of the model training data. These attacks may pose significant privacy risks to individuals whose sensitive data were used for training, which motivates the use of defenses such as differential privacy, often at the cost of high accuracy losses. MIAs exploit the differences in the behavior of a model when making predictions on samples it has seen during training (members) versus those it has not seen (non-members). Several studies have pointed out that model overfitting is the major factor contributing to these differences in behavior and, consequently, to the success of MIAs. However, the literature also shows that even non-overfitted ML models can leak information about a small subset of their training data. In this paper, we investigate the root causes of membership inference vulnerabilities beyond traditional overfitting concerns and suggest targeted defenses. We empirically analyze the characteristics of the training data samples vulnerable to MIAs in models that are not overfitted (and hence able to generalize). Our findings reveal that these samples are often outliers within their classes (e.g., noisy or hard to classify). We then propose potential defensive strategies to protect these vulnerable samples and enhance the privacy-preserving capabilities of ML models. Our code is available at https://github.com/najeebjebreel/mia_analysis.
In several jurisdictions, the regulatory framework on the release and sharing of personal data is being extended to machine learning (ML). The implicit assumption is that disclosing a trained ML model entails a privacy risk for any personal data used in training comparable to directly releasing those data. However, given a trained model, it is necessary to mount a privacy attack to make inferences on the training data. In this concept paper, we examine the main families of privacy attacks against predictive and generative ML, including membership inference attacks (MIAs), property inference attacks, and reconstruction attacks. Our discussion shows that most of these attacks seem less effective in the real world than what a prima face interpretation of the related literature could suggest.
Privacy models were introduced in privacy-preserving data publishing and statistical disclosure control with the promise to end the need for costly empirical assessment of disclosure risk. We examine how well this promise is kept by the main privacy models. We find they may fail to provide adequate protection guarantees because of problems in their definition or incur unacceptable trade-offs between privacy protection and utility preservation. Specifically, k-anonymity may not entirely exclude disclosure if enforced with deterministic mechanisms or without constraints on the confidential values. On the other hand, differential privacy (DP) incurs unacceptable utility loss for small budgets and its privacy guarantee becomes meaningless for large budgets. In the latter case, an ex post empirical assessment of disclosure risk becomes necessary, undermining the main appeal of privacy models. Whereas the utility preservation of DP can only be improved by relaxing its privacy guarantees, we argue that a semantic reformulation of k-anonymity can offer more robust privacy without losing utility with respect to traditional syntactic k-anonymity.
Large language models (LLMs) have recently revolutionized language processing tasks but have also brought ethical and legal issues. LLMs have a tendency to memorize potentially private or copyrighted information present in the training data, which might then be delivered to end users at inference time. When this happens, a naive solution is to retrain the model from scratch after excluding the undesired data. Although this guarantees that the target data have been forgotten, it is also prohibitively expensive for LLMs. Approximate unlearning offers a more efficient alternative, as it consists of ex post modifications of the trained model itself to prevent undesirable results, but it lacks forgetting guarantees because it relies solely on empirical evidence. In this work, we present DP2Unlearning, a novel LLM unlearning framework that offers formal forgetting guarantees at a significantly lower cost than retraining from scratch on the data to be retained. DP2Unlearning involves training LLMs on textual data protected using ε-differential privacy (DP), which later enables efficient unlearning with the guarantees against disclosure associated with the chosen ε. Our experiments demonstrate that DP2Unlearning achieves similar model performance post-unlearning, compared to an LLM retraining from scratch on retained data -- the gold standard exact unlearning -- but at approximately half the unlearning cost. In addition, with a reasonable computational cost, it outperforms approximate unlearning methods at both preserving the utility of the model post-unlearning and effectively forgetting the targeted information.
Human facial data offers valuable potential for tackling classification problems, including face recognition, age estimation, gender identification, emotion analysis, and race classification. However, recent privacy regulations, particularly the EU General Data Protection Regulation, have restricted the collection and usage of human images in research. As a result, several previously published face data sets have been removed from the internet due to inadequate data collection methods and privacy concerns. While synthetic data sets have been suggested as an alternative, they fall short of accurately representing the real data distribution. Additionally, most existing data sets are labeled for just a single task, which limits their versatility. To address these limitations, we introduce the Multi-Task Face (MTF) data set, designed for various tasks, including face recognition and classification by race, gender, and age, as well as for aiding in training generative networks. The MTF data set comes in two versions: a non-curated set containing 132,816 images of 640 individuals and a manually curated set with 5,246 images of 240 individuals, meticulously selected to maximize their classification quality. Both data sets were ethically sourced, using publicly available celebrity images in full compliance with copyright regulations. Along with providing detailed descriptions of data collection and processing, we evaluated the effectiveness of the MTF data set in training five deep learning models across the aforementioned classification tasks, achieving up to 98.88\% accuracy for gender classification, 95.77\% for race classification, 97.60\% for age classification, and 79.87\% for face recognition with the ConvNeXT model. Both MTF data sets can be accessed through the following link. https://github.com/RamiHaf/MTF_data_set
Synthetic data generation is one approach for sharing individual-level data. However, to meet legislative requirements, it is necessary to demonstrate that the individuals' privacy is adequately protected. There is no consolidated standard for measuring privacy in synthetic data. Through an expert panel and consensus process, we developed a framework for evaluating privacy in synthetic data. Our findings indicate that current similarity metrics fail to measure identity disclosure, and their use is discouraged. For differentially private synthetic data, a privacy budget other than close to zero was not considered interpretable. There was consensus on the importance of membership and attribute disclosure, both of which involve inferring personal information about an individual without necessarily revealing their identity. The resultant framework provides precise recommendations for metrics that address these types of disclosures effectively. Our findings further present specific opportunities for future research that can help with widespread adoption of synthetic data.
The threat of reconstruction attacks has led the U.S. Census Bureau (USCB) to replace in the Decennial Census 2020 the traditional statistical disclosure limitation based on rank swapping with one based on differential privacy (DP), leading to substantial accuracy loss of released statistics. Yet, it has been argued that, if many different reconstructions are compatible with the released statistics, most of them do not correspond to actual original data, which protects against respondent reidentification. Recently, a new attack has been proposed, which incorporates the confidence that a reconstructed record was in the original data. The alleged risk of disclosure entailed by such confidence-ranked reconstruction has renewed the interest of the USCB to use DP-based solutions. To forestall a potential accuracy loss in future releases, we show that the proposed reconstruction is neither effective as a reconstruction method nor conducive to disclosure as claimed by its authors. Specifically, we report empirical results showing the proposed ranking cannot guide reidentification or attribute disclosure attacks, and hence fails to warrant the utility sacrifice entailed by the use of DP to release census statistical data.
The objective of digital forgetting is, given a model with undesirable knowledge or behavior, obtain a new model where the detected issues are no longer present. The motivations for forgetting include privacy protection, copyright protection, elimination of biases and discrimination, and prevention of harmful content generation. Effective digital forgetting has to be effective (meaning how well the new model has forgotten the undesired knowledge/behavior), retain the performance of the original model on the desirable tasks, and be scalable (in particular forgetting has to be more efficient than retraining from scratch on just the tasks/data to be retained). This survey focuses on forgetting in large language models (LLMs). We first provide background on LLMs, including their components, the types of LLMs, and their usual training pipeline. Second, we describe the motivations, types, and desired properties of digital forgetting. Third, we introduce the approaches to digital forgetting in LLMs, among which unlearning methodologies stand out as the state of the art. Fourth, we provide a detailed taxonomy of machine unlearning methods for LLMs, and we survey and compare current approaches. Fifth, we detail datasets, models and metrics used for the evaluation of forgetting, retaining and runtime. Sixth, we discuss challenges in the area. Finally, we provide some concluding remarks.
$ε$-Differential privacy (DP) is a well-known privacy model that offers strong privacy guarantees. However, when applied to data releases, DP significantly deteriorates the analytical utility of the protected outcomes. To keep data utility at reasonable levels, practical applications of DP to data releases have used weak privacy parameters (large $ε$), which dilute the privacy guarantees of DP. In this work, we tackle this issue by using an alternative formulation of the DP privacy guarantees, named $ε$-individual differential privacy (iDP), which causes less data distortion while providing the same protection as DP to subjects. We enforce iDP in data releases by relying on attribute masking plus a pre-processing step based on data microaggregation. The goal of this step is to reduce the sensitivity to record changes, which determines the amount of noise required to enforce iDP (and DP). Specifically, we propose data microaggregation strategies designed for iDP whose sensitivities are significantly lower than those used in DP. As a result, we obtain iDP-protected data with significantly better utility than with DP. We report on experiments that show how our approach can provide strong privacy (small $ε$) while yielding protected data that do not significantly degrade the accuracy of secondary data analysis.
Training deep neural networks (DNNs) usually requires massive training data and computational resources. Users who cannot afford this may prefer to outsource training to a third party or resort to publicly available pre-trained models. Unfortunately, doing so facilitates a new training-time attack (i.e., backdoor attack) against DNNs. This attack aims to induce misclassification of input samples containing adversary-specified trigger patterns. In this paper, we first conduct a layer-wise feature analysis of poisoned and benign samples from the target class. We find out that the feature difference between benign and poisoned samples tends to be maximum at a critical layer, which is not always the one typically used in existing defenses, namely the layer before fully-connected layers. We also demonstrate how to locate this critical layer based on the behaviors of benign samples. We then propose a simple yet effective method to filter poisoned samples by analyzing the feature differences between suspicious and benign samples at the critical layer. We conduct extensive experiments on two benchmark datasets, which confirm the effectiveness of our defense.
In recent years, it has been claimed that releasing accurate statistical information on a database is likely to allow its complete reconstruction. Differential privacy has been suggested as the appropriate methodology to prevent these attacks. These claims have recently been taken very seriously by the U.S. Census Bureau and led them to adopt differential privacy for releasing U.S. Census data. This in turn has caused consternation among users of the Census data due to the lack of accuracy of the protected outputs. It has also brought legal action against the U.S. Department of Commerce. In this paper, we trace the origins of the claim that releasing information on a database automatically makes it vulnerable to being exposed by reconstruction attacks and we show that this claim is, in fact, incorrect. We also show that reconstruction can be averted by properly using traditional statistical disclosure control (SDC) techniques. We further show that the geographic level at which exact counts are released is even more relevant to protection than the actual SDC method employed. Finally, we caution against confusing reconstruction and reidentification: using the quality of reconstruction as a metric of reidentification results in exaggerated reidentification risk figures.
In federated learning (FL), a set of participants share updates computed on their local data with an aggregator server that combines updates into a global model. However, reconciling accuracy with privacy and security is a challenge to FL. On the one hand, good updates sent by honest participants may reveal their private local information, whereas poisoned updates sent by malicious participants may compromise the model's availability and/or integrity. On the other hand, enhancing privacy via update distortion damages accuracy, whereas doing so via update aggregation damages security because it does not allow the server to filter out individual poisoned updates. To tackle the accuracy-privacy-security conflict, we propose {\em fragmented federated learning} (FFL), in which participants randomly exchange and mix fragments of their updates before sending them to the server. To achieve privacy, we design a lightweight protocol that allows participants to privately exchange and mix encrypted fragments of their updates so that the server can neither obtain individual updates nor link them to their originators. To achieve security, we design a reputation-based defense tailored for FFL that builds trust in participants and their mixed updates based on the quality of the fragments they exchange and the mixed updates they send. Since the exchanged fragments' parameters keep their original coordinates and attackers can be neutralized, the server can correctly reconstruct a global model from the received mixed updates without accuracy loss. Experiments on four real data sets show that FFL can prevent semi-honest servers from mounting privacy attacks, can effectively counter poisoning attacks and can keep the accuracy of the global model.
TREs are widely, and increasingly used to support statistical analysis of sensitive data across a range of sectors (e.g., health, police, tax and education) as they enable secure and transparent research whilst protecting data confidentiality. There is an increasing desire from academia and industry to train AI models in TREs. The field of AI is developing quickly with applications including spotting human errors, streamlining processes, task automation and decision support. These complex AI models require more information to describe and reproduce, increasing the possibility that sensitive personal data can be inferred from such descriptions. TREs do not have mature processes and controls against these risks. This is a complex topic, and it is unreasonable to expect all TREs to be aware of all risks or that TRE researchers have addressed these risks in AI-specific training. GRAIMATTER has developed a draft set of usable recommendations for TREs to guard against the additional risks when disclosing trained AI models from TREs. The development of these recommendations has been funded by the GRAIMATTER UKRI DARE UK sprint research project. This version of our recommendations was published at the end of the project in September 2022. During the course of the project, we have identified many areas for future investigations to expand and test these recommendations in practice. Therefore, we expect that this document will evolve over time.
We introduce a new privacy model relying on bistochastic matrices, that is, matrices whose components are nonnegative and sum to 1 both row-wise and column-wise. This class of matrices is used to both define privacy guarantees and a tool to apply protection on a data set. The bistochasticity assumption happens to connect several fields of the privacy literature, including the two most popular models, k-anonymity and differential privacy. Moreover, it establishes a bridge with information theory, which simplifies the thorny issue of evaluating the utility of a protected data set. Bistochastic privacy also clarifies the trade-off between protection and utility by using bits, which can be viewed as a natural currency to comprehend and operationalize this trade-off, in the same way than bits are used in information theory to capture uncertainty. A discussion on the suitable parameterization of bistochastic matrices to achieve the privacy guarantees of this new model is also provided.