Search

Traditional centralized federated learning relies on a trusted central server for model aggregation, creating a vulnerability to single-point failures. In contrast, existing decentralized federated learning schemes elect a node temporarily in each iteration cycle to aggregate the model, but cannot ensure the complete trustworthiness of the elected node. To solve the aforementioned issues, this paper proposed a blockchain-based federated learning model aggregation approach that assigned the task of model aggregation to numerous miners instead of a single node. Miners proposed various candidate aggregation solutions and generated corresponding blocks, then it determined the main chain based on the highest accuracy chain principle to achieve consensus among nodes. Additionally, to counteract malicious training nodes, it introduced a training node selection mechanism based on staking training coins, allowing nodes to participate in training by staking training coins, with the system rewarding or penalizing them based on their contribution to the model. Simulation results demonstrate that with 10%, 20% and 30% malicious nodes in the system, the accuracy of this approach is respectively 8.64, 19.89, and 22.93 percent point higher than that of the federated averaging (FedAvg) scheme, and it also performs well in non-IID data training scenarios. In conclusion, this approach enhances the credibility of the federated learning aggregation process and ensures the effectiveness of federated learning training. [ABSTRACT FROM AUTHOR]