A blockchain-based privacy-preserving reputation consensus federated learning

To address the challenges of privacy leakage and malicious node attacks in traditional federated learning, this study proposes BPRFL (Blockchain-based Privacy-preserving Reputation Consensus Federated Learning), a novel architecture integrating blockchain technology with federated learning. First, a noise-separated differential privacy mechanism is introduced, where noise generation is architecturally decoupled from local clients through VRF-based noise group formation. Unlike existing approaches where clients self-generate noise (e.g., PPFLChain, Biscotti), this prevents malicious participants from crafting adversarial noise to camouflage poisoned updates. Second, a dual-constraint reputation consensus scheme is proposed, employing confidence interval consistency verification to detect collusion attacks. This extends beyond performance-based reputation systems (e.g., BFLC) by evaluating both model quality and evaluation consistency, preventing mutual election among colluding nodes. Third, a dynamic committee scaling mechanism adaptively adjusts committee size based on cumulative reputation scores, improving consensus efficiency by 20% compared to fixed-size committees while maintaining Byzantine fault tolerance. Experimental results demonstrate that, under a privacy budget of 0.3, the differential privacy mechanism in BPRFL effectively protects on-chain model parameters from inference attacks. Under 50% malicious clients performing label-flipping attacks, BPRFL achieves 5.98% higher accuracy than traditional FL and 2.09% higher than BFLC, while maintaining superior consensus efficiency validated across different model architectures and real-world datasets.