FedPG : a privacy-friendly and universal method for solving non-IID data in federated learning

Xue, B. and Zhang, J. and Chen, B. and Meng, W. (2025) FedPG : a privacy-friendly and universal method for solving non-IID data in federated learning. Pattern Analysis and Applications, 28 (2): 100.

Text (sn-article)
sn-article.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (1MB)

Abstract

Federated learning (FL) is a privacy-preserving distributed learning framework which could harness the potential of decentralized multimedia data. However, a significant hurdle lies in the non-uniform distribution of data among clients, leading to slow convergence and subpar accuracy in the global model. Although several approaches have been proposed to address this challenge, two key limitations remain. First, these methods frequently require access to information about local data or even the raw data, which raises significant privacy concerns for clients. Second, these methods struggle to perform well in a common non-IID scenario: class missingness, and they often fail to fully resolve the issue of client drift. In response, in this paper, we propose a privacy-friendly and universal method FedPG to solve non-IID data in FL. The core idea behind FedPG is to leverage homogeneous virtual data to alleviate both data heterogeneity and client drift. Specifically, FedPG introduces a novel image generation method based on prototype loss, which does not require any additional privacy-sensitive information. This approach generates synthetic datasets aligned with the global distribution to effectively assist local training. Besides, we also design a local training method that is suitable for scenarios involving class missingness, enabling both feature adaptation and classifier de-biasing. The comprehensive experiments demonstrate the efficacy of our FedPG framework. In the majority of cases, FedPG not only achieves superior accuracy but also exhibits accelerated convergence rates compared to alternative approaches.