Committee Consensus-based Auditable Privacy-preserving Multi-round Federated Learning Aggregation

Federated Learning (FL) faces significant privacy, security, efficiency, and auditability challenges in multi-round training. This paper introduces CC-APPLF, a Committee Consensus-based Auditable Privacy-Preserving Multi-round FL aggregation framework. It innovatively integrates blockchain and advanced cryptography to create a decentralized, auditable, and privacy-centric solution. Key mechanisms include committee-based aggregation with separated decryption responsibilities, efficient dual-masking for update privacy, and optimizations for multi-round efficiency, all coordinated and recorded immutably via blockchain. CC-APPLF addresses traditional FL limitations such as centralization, opacity, and inefficiency. Experimental results demonstrate comparable accuracy to FedAvg [1], significantly improved communication efficiency over naive encryption, and competitive training times against existing secure schemes. It offers a robust solution for deploying secure and efficient distributed intelligent applications.