Enhancing Privacy-Preserving Federated Learning: A Robust and Resilient Approach

Federated learning has garnered considerable attention for its capacity to aggregate efficient and accurate models while preserving user’s data privacy. Nonetheless, conventional federated learning frameworks remain susceptible to malicious threats like collusion and inference attacks. We design a robust and resilient enhancing privacy-preserving federated learning framework based on blockchain (EPFL) to address these threats. Initially, we meticulously design a two-round encryption scheme that allows users to upload encryption gradients while the server can still perform aggregation. Subsequently, a secure communication protocol termed the “server-blockchain-user” is devised based on blockchain architecture. This protocol guarantees the openness, transparency, and traceability of the entire training process, while also offering resilience in users’ drop-out and joining that may arise during the aggregation process. Finally, experimental assessments are conducted to appraise the robustness and accuracy of EPFL. The empirical findings demonstrate that EPFL not only demonstrates robustness against collusion and inference attacks, but also adeptly handles user exits and entries without compromising the confidentiality of user keys. Moreover, EPFL attains model accuracy commensurate with the FedAvg.