pFedCE: Personalized Federated Learning Based on Contribution Evaluation
Quanhong Li, Xuehui Li, Zelei Liu, Hongxing Qi · 2024
Federated learning (FL) enables collaborative model training across clients without sharing raw data, ensuring privacy while utilizing decentralized data. However, the heterogeneous distribution of data among clients can significantly degrade the performance of the global model. To address this problem, we propose a personalized FL (pFL) framework pFedCE. Unlike traditional FL, pFedCE introduces a Shapley value-based contribution evaluation during model aggregation, adaptively adjusting aggregating weights based on contribution to generate a personalized aggregation model for each client. Since direct access to client data for contribution evaluation would violate principles of FL, we utilize a Generative Adversarial Network (GAN) to generate synthetic data on the server that preserve the statistical properties of client data. We evaluated pFedCE on three benchmark datasets in three types of data distribution. The results show that pFedCE outperforms six other pFL methods.