VDFChain: Secure and verifiable decentralized federated learning via committee-based blockchain

Abstract

Decentralized federated learning tries to address the single point of failure and privacy issue of federated learning by leveraging committee-based blockchain, which has been extensively studied among academic and industrial fields. The introduction of committees improves the efficiency of decentralized federated learning. However, it also is prone to attacks from Byzantine committee members, which interfere with the correctness of the global model by modifying aggregation results. Therefore, the security of committees is the key challenge for decentralized federated learning via committee-based blockchain. To solve this problem, in this paper, we propose VDFChain, a secure and verifiable decentralized federated learning scheme via committee-based blockchain. Specifically, based on the polynomial commitment technique, we propose a trusted committee mechanism, which can defend against attacks from Byzantine committee members and ensure the correctness of the aggregation model. Moreover, we use lossless masking techniques and committee mechanisms to effectively provide secure aggregation. For Byzantine attacks in decentralized federated learning, different from traditional defense methods against it, the VDFChain improves the fault tolerance of decentralized federated learning and provides a feasible and practical solution to build a secure decentralized federated learning. Security analysis shows that our scheme is provably secure. We have conducted extensive comparison experiments to evaluate the performance of the proposed framework, and experimental results show that our scheme has superior computational and communication performance compared to the state-of-the-art schemes.