The Convergence of Client Heterogeneity and Generative Models in Personalized Federated Learning

[Submitted on 23 Aug 2023]

Download a PDF of the paper titled “PFL-GAN: When Client Heterogeneity Meets Generative Models in Personalized Federated Learning,” authored by Achintha Wijesinghe and 2 others.

Download PDF

Abstract: Recent advancements in generative learning models have led to a growing interest in federated learning (FL) based on generative adversarial network (GAN) models. GANs can capture the underlying client data structure in FL and generate samples that resemble the original data distribution without compromising the privacy of raw data. While most existing GAN-based FL works focus on training a global model, Personalized FL (PFL) can sometimes be more effective due to client data heterogeneity in terms of distinct data sample distributions, feature spaces, and labels. To address client heterogeneity in GAN-based FL, we propose a novel GAN sharing and aggregation strategy for PFL called PFL-GAN. PFL-GAN learns the similarity among clients and develops a weighted collaborative data aggregation. Empirical results from rigorous experimentation on several well-known datasets demonstrate the effectiveness of PFL-GAN.