A multi-center federated learning mechanism based on consortium blockchain for data secure sharing

Currently, federated learning enables hospitals to collaborate on model training without disclosing patient privacy data. However, it still faces challenges such as single point of failure and communication inefficiency. For this reason, this study innovatively combines consortium blockchain and federated learning. A multicenter federated learning mechanism based on consortium blockchain (MCFLM-CB) is proposed to optimize the security and efficiency of data collaboration and sharing. Firstly, the MCFLM-CB model uses the multi-party co-management feature of the consortium blockchain to replace the central server of federated learning, so that the system performs the training of the federated learning model in multiple centers. It also achieves the elimination of the disadvantages that a single centralized server controls the data model. Secondly, we propose a Dynamic Grouping-based Practical Byzantine Fault Tolerant (DG-PBFT) consensus algorithm. The algorithm performs regrouping and center node selection based on node state changes. It improves the consensus algorithm in blockchain system adaptive ability. Finally, we propose a reputation value-based weighted federal average algorithm. By synthesizing multiple reputation attributes to evaluate the reputation of participants, it comprehensively reflects the node performance. The accuracy and reliability of reputation values are improved. To prove the effectiveness of the method, we validated it on 12 large-scale standardized biomedical image sets MedMNIST. The results show that the model achieves 93.2% accuracy and significantly improves the efficiency of the blockchain.