Privacy-Preservation Robust Federated Learning with Blockchain-based Hierarchical Framework

Federated Learning (FL) is a learning architecture in which multiple clients use local data to train gradients and submit them to a server for global aggregation. However, achieving reliable federated learning in untrusted environments is challenging. Malicious users and servers can perform Byzantine attacks to manipulate the global model, and curious servers can also extract user privacy from update gradients. Privacy-preserving methods usually consume many communication and computing resources and are difficult to combine with robust aggregation mechanisms. In this paper, we propose an anomaly gradient detection approach based on cosine distance detection using Shamir secret sharing, which can detect anomalous gradients without exposing the real gradients. Furthermore, we use the blockchain to build a decentralized hierarchical FL framework, eliminating the risk of malicious servers and reducing communication overhead. The experimental results show that our approach can protect user privacy and the reliability of aggregation results simultaneously, and the layered framework can effectively reduce communication overhead.