Personalized Federated Learning for Statistical Heterogeneity
Muhammad Firdaus, Kyung-Hyune Rhee
TL;DR
Federated learning promises privacy-preserving collaboration but suffers from statistical heterogeneity due to non-IID client data, leading to poor convergence and weak personalization. The paper surveys the current landscape of personalized federated learning, categorizing methods into data-based, model-based, and similarity-based approaches, and discusses practical challenges such as privacy, trust, and benchmarking. It presents a concise objective for PFL and outlines a taxonomy, key techniques, and representative works, including FedAvg variants, meta-learning, transfer learning, regularization, distillation, and clustering-based methods. The analysis highlights the need for privacy-preserving, trustworthy PFL and the importance of diverse datasets and benchmarks to drive progress. The work serves as a roadmap for researchers and practitioners aiming to deploy effective PFL systems in privacy-sensitive settings.
Abstract
The popularity of federated learning (FL) is on the rise, along with growing concerns about data privacy in artificial intelligence applications. FL facilitates collaborative multi-party model learning while simultaneously ensuring the preservation of data confidentiality. Nevertheless, the problem of statistical heterogeneity caused by the presence of diverse client data distributions gives rise to certain challenges, such as inadequate personalization and slow convergence. In order to address the above issues, this paper offers a brief summary of the current research progress in the field of personalized federated learning (PFL). It outlines the PFL concept, examines related techniques, and highlights current endeavors. Furthermore, this paper also discusses potential further research and obstacles associated with PFL.