FedG3FA: Three-Stage GAN-Aided Target Feature Alignment for Secure Data Sharing in Federated Learning System

Abstract

Federated learning (FL) allows distributed clients to train model collaboratively without sharing the original data. However, using private model updates often makes traditional FL systems susceptible to privacy leakage problem. In addition, the performance of existing FL methods is often limited by statistical heterogeneity problem. In order to solve both privacy leakage and statistical heterogeneity problems, we propose a three-stage targeted feature alignment FL framework named FedG3FA. In the first stage, each client trains a generator through generative adversarial training and the generator will be utilized for data interaction instead of private model. After that, in the second stage, the generators will be aligned by our proposed Domain Pulling Network and then aggregated to a global one. Finally, in the third stage, the global generator will be used to train the private model for each client. The effectiveness of our method is verified on medical care and computer vision scenarios including five datasets. The experimental results suggest that our method not only achieves a high level of privacy protection performance, but also remains competitive classification accuracy.