The security of Fully Homomorphic Encryption (FHE) has received a lot of attention in recent years with new security notions emerging to better understand the practical attacks that may threaten the real-world deployments of passively secure FHE schemes. One such new notions is CPAD a slight extension of CPA security modelling a passive adversary who is granted access to a decryption oracle accepting only well-formed ciphertexts. While successful CPAD attacks have initially been performed on approximate FHE schemes such as CKKS, recent works have also demonstrated practical CPAD attacks on all mainstream non-approximate FHE, such as BFV, BGV or TFHE. Despite their clear computational practicality, these latter attacks however focus on the abstract security game defining CPAD security. In this paper, we show how to concretely build on these to mount successful FHE key recovery attacks in the Federated Learning (FL) setting, an application scenario of choice for FHE techniques. In FL, participating entities or workers encrypt successive model updates based on their local training data, enabling a central server to aggregate them in order to homomorphically update a global model. As this paper demonstrates, this environment provides a playground for an attacker to launch key recovery attacks against the FHE underlying the secure aggregation mechanism. As such, our findings reveal substantial stealthy key-recovery threats from both the server and a single worker, with very limited impact on the FL training progression or final model quality.

Machine learning (ML) algorithms are heavily based on the availability of training data, which, depending on the domain, often includes sensitive information about data providers. This raises critical privacy concerns. Anonymization techniques have emerged as a practical solution to address these issues by generalizing features or suppressing data to make it more difficult to accurately identify individuals. Although recent studies have shown that privacy-enhancing technologies can influence ML predictions across different subgroups, thus affecting fair decision-making, the specific effects of anonymization techniques, such as k-anonymity, ℓ-diversity, and t-closeness, on ML fairness remain largely unexplored. In this work, we systematically audit the impact of anonymization techniques on ML fairness, evaluating both individual and group fairness. Our quantitative study reveals that anonymization can degrade group fairness metrics by up to fourfold. Conversely, similarity-based individual fairness metrics tend to improve under stronger anonymization, largely as a result of increased input homogeneity. By analyzing varying levels of anonymization across diverse privacy settings and data distributions, this study provides critical insights into the trade-offs between privacy, fairness, and utility, offering actionable guidelines for responsible AI development. Our code is publicly available at: https://github.com/hharcolezi/anonymity-impact-fairness.

Algorithmic fairness is a critical challenge in building trustworthy Machine Learning (ML) models. ML classifiers strive to make predictions that closely match real-world observations (ground truth). However, if the ground truth data itself reflects biases against certain sub-populations, a dilemma arises: prioritize fairness and potentially reduce accuracy, or emphasize accuracy at the expense of fairness. This work proposes a novel training framework that goes beyond achieving high accuracy. Our framework trains a classifier to not only deliver optimal predictions but also to identify potential fairness risks associated with each prediction. To do so, we specify a dual-labeling strategy where the second label contains a per-prediction fairness evaluation, referred to as an unfairness risk evaluation. In addition, we identify a subset of samples as highly vulnerable to group-unfair classifiers. Our experiments demonstrate that our classifiers attain optimal accuracy levels on both the Adult-Census-Income and Compas-Recidivism datasets. Moreover, they identify unfair predictions with nearly 75% accuracy at the cost of expanding the size of the classifier by 45%.

Federated Learning (FL) enables the distributed training of a model across multiple data owners under the orchestration of a central server responsible for aggregating the models generated by the different clients. However, the original approach of FL has significant shortcomings related to privacy and fairness requirements. Specifically, the observation of the model updates may lead to privacy issues, such as membership inference attacks, while the use of imbalanced local datasets can introduce or amplify classification biases, especially for minority groups. In this work, we show that these biases can be exploited to increase the likelihood of privacy attacks against these groups. To do so, we propose a novel inference attack exploiting the knowledge of group fairness metrics during the training of the global model. Then to thwart this attack, we define a fairness-aware encrypted-domain aggregation algorithm that is differentially-private by design thanks to the approximate precision loss of the threshold multi-key CKKS homomorphic encryption scheme. Finally, we demonstrate the good performance of our proposal both in terms of fairness and privacy through experiments conducted over three real datasets.

Algorithmic fairness is a critical challenge in building trustworthy Machine Learning (ML) models. ML classifiers strive to make predictions that closely match real-world observations (ground truth). However, if the ground truth data itself reflects biases against certain sub-populations, a dilemma arises: prioritize fairness and potentially reduce accuracy, or emphasize accuracy at the expense of fairness. This work proposes a novel training framework that goes beyond achieving high accuracy. Our framework trains a classifier to not only deliver optimal predictions but also to identify potential fairness risks associated with each prediction. To do so, we specify a dual-labeling strategy where the second label contains a per-prediction fairness evaluation, referred to as an unfairness risk evaluation. In addition, we identify a subset of samples as highly vulnerable to group-unfair classifiers. Our experiments demonstrate that our classifiers attain optimal accuracy levels on both the Adult-Census-Income and Compas-Recidivism datasets. Moreover, they identify unfair predictions with nearly 75% accuracy at the cost of expanding the size of the classifier by a mere 45%.