Privacy-Preserving Clustered Peer-to-Peer Learning

In peer-to-peer federated learning (FL), clients periodically exchange their training models directly. This raises a privacy concern with respect to peers. In this paper, we propose an algorithm for peer-to-peer federated learning that is fully decentralized, that can work in FL systems with a large number of clients via efficient clustering and that is privacy-preserving. Our proposed algorithm aggregates models from clients in the same cluster to reduce communication overhead and improve learning accuracy. In each iteration of model aggregation, the algorithm uses Shamir's secret sharing scheme to preserve clients' privacy. We provide the convergence analysis of the algorithm. We provide experimental results to assess the performance of PCPL compared to some existing baselines. The analytical and experimental results demonstrate that our proposed algorithm is outperforms the state-of-the-art in privacy preserving FL algorithms.