Beyond Random Selection: A Perspective from Model Inversion in Personalized Federated Learning

Abstract

With increasing concern for privacy issues in data, federated learning has emerged as one of the most prevalent approaches to collaboratively train statistical models without disclosing raw data. However, heterogeneity among clients in federated learning hinders optimization convergence and generalization performance. For example, clients usually differ in data distributions, network conditions, input/output dimensions, and model architectures, leading to the misalignment of clients’ participation in training and degrading the model performance. In this work, we propose PFedRe, a personalized approach that introduces individual relevance , measured by Wasserstein distances among dummy datasets, into client selection in federated learning. The server generates dummy datasets from the inversion of local model updates, identifies clients with large distribution divergences, and aggregates updates from high relevant clients. Theoretically, we perform a convergence analysis of PFedRe and quantify how selection affects the convergence rate. We empirically demonstrate the efficacy of our framework on a variety of non-IID datasets. The results show that PFedRe outperforms other client selection baselines in the context of heterogeneous settings.