Robust peer-to-peer learning via secure multi-party computation

Abstract

To solve the data island problem, federated learning (FL) provides a solution paradigm where each client sends the model parameters but not the data to a server for model aggregation. Peer-to-peer (P2P) federated learning further improves the robustness of the system, in which there is no server and each client communicates directly with the other. For secure aggregation, secure multi-party computing (SMPC) protocols have been utilized in peer-to-peer manner. However, the ideal SMPC protocols could fail when some clients drop out. In this paper, we propose a robust peer-to-peer learning (RP2PL) algorithm via SMPC to resist clients dropping out. We improve the segment-based SMPC protocol by adding a check and designing the generation method of random segments. In RP2PL, each client aggregates their models by the improved robust secure multi-part computation protocol when finishes the local training. Experimental results demonstrate that the RP2PL paradigm can mitigate clients dropping out with no significant degradation in performance.