Mitigating Membership Inference Attacks in Federated Learning
Résumé
Federated Learning (FL) is a machine learning technique that allows multiple data owners to collaborate in training a model without sharing their training data. However, FL systems are still vulnerable to privacy attacks, where malicious participants can infer information about other participants' data by using the exchanged parameters. Membership inference attacks are a type of privacy attack that allows an attacker to infer whether a data sample was used by a target participant to train its own local model, thereby threatening data privacy. To address this issue, the paper proposes PASTEL, a novel FL privacy-preserving mechanism based on internal generalization gap minimization. The experimental evaluation of PASTEL shows that it reduces the membership inference attack success rate closely to 50% (best-case scenario) with a negligible impact on local models' utility.