Security and Verifiability in Federated Learning: A Zero-Knowledge Reputation-Based Blockchain Framework

Abstract

Federated Learning (FL) enables collaborative training without centralizing sensitive data but faces challenges, including client authenticity, verifiable training participation, and secure aggregation. To overcome these challenges, we propose a novel framework, Zero-Knowledge Reputation-aware Blockchain Federated Learning (ZK-RBFL), which integrates blockchain, FL, Homomorphic Encryption (HE), and zero-knowledge proofs (ZKP). In the proposed ZK-RBFL framework, initially the clients undergo lightweight token-based authentication and then generate ZKP to provide cryptographic evidence of honest local training participation and reported inference accuracy before contributing their model updates. The model updates are encrypted using the CKKS HE mechanism to prevent any potential model inversion attacks. These encrypted model updates are stored on IPFS, with their corresponding CIDs recorded on the blockchain to ensure immutability. Further, ZK-RBFL enables mutual client verification of ZKPs to reduce server bottlenecks and enhance accountability. To ensure fairness and robustness in a distributed environment, we introduce a democratic blockchain consensus mechanism named Proof of Reputation-Weighted Voting (PoRWV) for block acceptance. Once consensus is reached, the encrypted model updates are aggregated using reputation-weighted averaging. We demonstrate the effectiveness of ZK-RBFL for brain tumor classification using a ZKP-compatible LeNet model for proof generation. Despite model simplicity, the global model achieves 94.22% accuracy. In addition, experiments with malicious clients and formal Scyther security analysis demonstrate that ZK-RBFL ensures both security and performance.