Client-based differential privacy federated learning

Le résumé fourni par la source

Deep learning provides better personalized services by training specific models through massive amounts of data. However, due to the problem of gradient leakage during model training, the original data uploaded by the users is restored and privacy leakage occurs. In order to prevent data leakage, this paper introduces a federated learning method to deal with the privacy issues brought by multi-user joint modeling. Gradients generated by the user’s local model training are uploaded to the aggregation server without being trained directly using the original user data. Under such a framework setting, the users’ original data still has a certain risk of being leaked. In order to strengthen the protection of users’ privacy, the training process is encrypted by combining the differential privacy mechanism and the federated learning system. The model parameters are stochastic to ensure that they cannot be acquired by adversaries. By adding Gaussian mechanism and Laplace mechanism, the influence of differential privacy on the convergence of federated learning model is analyzed. The Laplace mechanism is a strict definition of differential privacy, while the Gaussian mechanism is a relaxed definition and allows adding less noise for privacy protection. The simulation results show that both mechanisms can achieve good model convergence effect and verify that differential privacy can produce better privacy protection effect with lower communication cost.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.