A Personalized Privacy-Preserving Scheme for Federated Learning

Federated learning (FL), which is a state-of-theart distributed machine learning (DML) model, brings the spare resources of mobile devices into full play and provides strong security guarantee to local sensitive data by local differential privacy. However, the introduction of noise data leads to an unignored reduction of model utility. In this paper, we consider the heterogeneity of privacy requirement for various participants in FL and propose a novel federated learning scheme (PGC-LDP) that lets users personally choose their privacy level based on federated stochastic gradient descent algorithm with local differential privacy. In the scheme, we design a new algorithm based on Nguyen's solution in client side and optimize aggregation method in server side. Moreover, we theoretically analyze the privacy guarantee and verify the utility of PGC-LDP on real-world dataset.