Fed2VAEs: An Efficient Privacy-Preserving Federated Learning Approach Based on Variational Autoencoders

Recently, federated learning (FL) has been threat-ened by the gradient inversion attack that infers user-private data from shared gradients. To cope with this problem, the differential privacy (DP) technique is widely employed in FL. However, when FL faces the non-independent identically distributed (non-IID) data scenarios, applying DP to protect user data privacy remains inefficient in terms of model accuracy and communication costs. In this paper, inspired by the Mixup data augmentation method, we propose a privacy-preserving FL approach called Fed2VAEs to address this problem. Specifically, we introduce a Mixup Module consisting of two variational autoencoders to remove the private information of user data. To balance the trade-off between data privacy and data utility, from the perspective of mutual information, a learning objective is proposed. We conduct extensive experiments under different non-IID data settings, and the experimental results show that Fed2VAEs can significantly reduce the communication cost and improve model accuracy (up to 8.57%) on the premise of successfully protecting user data privacy.