Blockchain-Enabled Federated Learning: A Dynamic-Grouping Privacy-Preserving Framework

With the mounting public concern for life safety, the rapid development of machine learning in the medical field has garnered significant attention. To address the privacy challenges associated with medical data, federated learning (FL) has emerged as a prevailing paradigm. However, despite its partial data security assurance, FL still falls short of meeting broader privacy and trust requirements in practical collaborative settings. In this paper, we propose a blockchain-assisted federated learning framework with dynamic grouping, aiming to reduce the direct exposure of individual client updates during collaborative training. By integrating Hyperledger Fabric with FL, the proposed framework combines dynamic grouping, group-based additive sharing, and temporary blinding in a permissioned environment. In the proposed protocol, users do not share raw training data, and only blinded share-level values together with aggregation-related records are involved in the on-chain process. We further provide a preliminary correctness and security discussion under the semi-honest setting and report proof-of-concept experimental results on MNIST. The current results indicate the feasibility of the proposed framework under the evaluated setting.