Privacy-Enhanced Federated Learning Algorithm Empowered by Blockchain and Game Theory

Federated learning enables participants to collaboratively train a global model through distributed training without sharing raw data. However, this distributed training is vulnerable to single-point failures, privacy leakage, and Byzantine attacks. To address these challenges simultaneously, we propose a privacy-enhanced federated learning algorithm empowered by blockchain and game theory (BGFL). First, we integrate blockchain technology into federated learning and establish a decentralized training paradigm, which effectively avoids the threat of single-point failures in centralized training. Second, we propose a privacy protection method that combines the Chinese Remainder Theorem and Shamir's Secret Sharing, providing dual privacy protection for participants in decentralized federated learning scenarios. Furthermore, leveraging the homomorphic properties of the proposed privacy-enhanced approach, a Byzantine attack defense mechanism is designed. Finally, a game-theoretic incentive mechanism is proposed to mitigate malicious behaviors during collaborative training, ensuring secure cooperation among all parties. Experimental results demonstrate that, compared to baseline methods, BGFL improves test accuracy by 0.93%–22.75%. Compared to other secret sharing-based federated learning schemes, BGFL requires only 5.91% of the computational overhead and 4.01% of the communication overhead at the same dataset scale. This enables efficient achievement of security and robustness objectives.