A Study on Trustworthy Aggregation Mechanisms for Blockchain-Based Federated Learning

Abstract

To investigate how blockchain technology can improve the reliability, speed and legality of federated learning, this paper concentrates on the process of model combination in federated learning. By dealing with the problems of single vulnerable points and insufficient trust in centralized combination, the paper comprehensively studies the technical methods of typical solutions like BlockFL, FedChain and LBFL, and develops a four-dimensional evaluation system including decentralization, system performance, security and robustness, and privacy disclosure. Three main conflicts have been found: the contradiction between the immutability on chain and the principle of data minimization; the conflict between high-level confidentiality protection and scarce resources on edge devices; and the potential centralization risk under apparent decentralization. According to these results, this paper puts forward a lightweight cooperative design concept named "on-chain evidence storage and off-chain computation", which serves as a theoretical basis and practical approach for trustworthy federated learning in very sensitive situations.