BE-DPFL: A Blockchain-Enhanced Privacy-Preserving Federated Learning Framework for Secure Edge Network Collaboration

Against the deep integration of digital transformation and AI, cross-institutional collaborative modeling hinges on efficient data circulation, yet data silos and privacy regulations hinder traditional centralized training. Federated Learning (FL) keeps data local but faces issues like weak centralized trust, inadequate privacy protection, and poor robustness in edge networks. Existing improvements, including via differential privacy (DP) and blockchain, among others, still suffer from centralized budget allocation, low consensus efficiency, or single-point-of-failure addressing, failing to jointly optimize trust, performance, and privacy. The limitations are exacerbated in high-frequency, resource-constrained edge environments. To tackle these challenges, this paper proposes BE-DPFL, a blockchain-enhanced differentially private FL framework that integrates on-chain trusted supervision and off-chain efficient training. It builds a lightweight blockchain trust layer with FL-PBFT consensus and smart contracts, introduces Random Projection–ADMM optimization, and designs a multi-objective adaptive gradient clipping/noise injection strategy. Experiments on CIFAR-10 and ChestX-ray14 demonstrate that BE-DPFL outperforms mainstream methods in consensus efficiency, communication overhead, privacy-accuracy balance, and robustness. It reduces communication costs by over 97%, achieves 100% privacy compliance, and maintains stable performance even under high disturbances. Ablation studies confirm the significant contributions of core components.