Ring-PPFL: A Ring Topology based Framework for Privacy-Preserving Federated Learning
Changji Wang, Boxuan Lin, Ning Liu, Zhen Liu, Jingjing Zhang, Qingqing Gan · 2024
Federated learning enables collaborative model training while preventing the centralization of data, thus ensuring data privacy. However, the potential leakage of model updates poses privacy risks. In this paper, we propose Ring-PPFL, a framework based on ring topology and double masks, to improve privacy preservation in federated learning. Specifically, Ring-PPFL adopts a ring communication mechanism to transmit local model parameters and uses a double-mask mechanism of two independent masks (one provided by the server and another by a trusted third party) in the communication to defend against collusion attacks. Experiments on MNIST and CIFAR-100 show that the model accuracy and convergence speed of Ring-PPFL are comparable to FedAvg, and the model accuracy is 10%-60% higher than differential privacy. At the same time, the GML value of Ring-PPFL is about 60 times higher than that of differential privacy using larger noise (1e-1); that is, the privacy preservation capability of Ring-PPFL is better than that of differential privacy. Our framework provides a feasible solution for deploying federated learning with enhanced privacy preservation.