PGPFL: A Parameter Guard-Based Efficient Personalized Federated Learning Framework with Local Differential Privacy

Data heterogeneity and privacy leakage pose inevitable challenges for federated learning in real-world scenarios. These obstacles have motivated the development of numerous personalized federated learning algorithms aimed at acquiring personalized models for individual clients. However, previous work in achieving personalization has typically proposed solutions from the perspective of the entire model or from the level of layer within the network structure. Diverging from them, we introduce a novel perspective, scrutinizing the personalization issue with finer granularity at the parameter level. In this paper, we propose PGPFL, a novel personalized federated learning framework incorporating differential privacy, which rapidly achieves the required local accuracy while providing privacy protection. PGPFL introduces configuring a guard for each parameter of the local network to constrain local updates for personalized learning, and calculate the importance value of each parameter of the current model during the local training process to determine the private parts. Additionally, PGPFL propose a valid client selection mechanism based on the response ratio and local accuracy to decrease single-round training time, enhancing training efficiency. Extensive experiments conducted under various simulated heterogeneous environments demonstrate that, in comparison with existing algorithms, our algorithm improves both average local accuracy and training efficiency.