An Optimized Scheme of Federated Learning Based on Differential Privacy

Abstract

With the arrival of the Web3 era, data has seen an explosive growth. The use of differential privacy mechanisms in federated learning has been proposed to protect user privacy and avoid security threats from data sharing. The core idea of this approach is that multiple clients train their local data, add noise to their client parameters, and then transmit them to a central server for parameter aggregation. However, there are still defects that need to be addressed. First, it is difficult to resist attacks from malicious clients, which means that user privacy is not fully protected. Second, it is challenging to add an appropriate amount of noise to achieve high model accuracy. Therefore, this paper proposes a bidirectional adaptive noise addition federated learning scheme, which adds adaptive noise satisfying the differential privacy mechanism to both the central server and clients to improve model accuracy. Considering the heterogeneity of client hardware, this paper samples gradients and samples separately to reduce communication costs and uses RMSprop to accelerate model training on both clients and central servers. Experimental results show that the proposed scheme enhances user privacy protection while maintaining high efficiency.