A Blockchain Federated Learning Scheme Based on Personalized Differential Privacy and Reputation Mechanisms

Federated learning is a distributed machine learning paradigm that effectively safeguards users' personal data from attackers. However, this approach is still susceptible to inference attacks that could compromise user privacy by analyzing gradient parameters during model training. To counteract this problem, differential privacy is applied to improve the privacy of federated learning. In this paper, we propose a blockchain-based federated learning model that leverages personalized differential privacy. Specifically, users can adjust levels of noise added to their local models based on their privacy requirements. The blockchain is integrated to avoid the central server's single-point-of-failure problem, thereby enhancing the security of the system overall. Furthermore, the proposed scheme incorporates reputation values to dynamically adjust aggregation weights, enabling faster model convergence and reducing the impact of malicious clients on the model. Experimental results demonstrate that the approach meets personalized privacy requirements while achieving high accuracy rates.