FedMCP: A Personalized Federated Learning Based on Kernel Distance Between Clients

Federated Learning (FL) is a distributed machine learning framework designed to safeguard data privacy. However, its performance and fairness are often hindered by the dual challenges of data heterogeneity (Non-IID) and privacy protection. This paper proposes a Federated Learning framework based on a Conditional Policy Network (CPN) and Differential Privacy (DP) to address these issues. The CPN dynamically generates weights for personalized and global features, balancing the trade-off between global consistency and local personalization, thereby mitigating the impact of data heterogeneity. Simultaneously, the differential privacy technique incorporates Gaussian noise into gradient updates to safeguard sensitive client data, minimizing the risk of privacy breaches. The framework's effectiveness was validated through experiments on multiple non-IID datasets. The results indicate that the proposed method surpasses mainstream approaches in global model performance and client fairness while striking a favorable balance between privacy protection and model performance. Additionally, this study examines the effects of noise intensity and gradient clipping thresholds on model performance, offering theoretical and practical insights for the design of more efficient Federated Learning frameworks.