FedBC: Blockchain-based Decentralized Federated Learning
Xin Wu, Zhi Wang, Jian Zhao, Yan Zhang, Yu Wu
Abstract
Federated learning enables participants to collaborate on model training without directly exchanging raw data. Existing federated learning methods often follow the parameter server architecture, using third-party collaborators to provide aggregation and key management. In this case, the central node obtains information uploaded by other nodes. Studies have shown that with this information, the central node can infer important information, which leads to data privacy leakage. In addition, the failure on the server node can also cause the entire system to fail. We designed a completely decentralized federated learning framework based on blockchain, thereby avoiding the privacy and failure risk of the centralized structure. Moreover, we develop the corresponding model training approach. Compared with the existing methods, our framework performs better in terms of accuracy, robustness, and privacy.