Communication-Efficient Personalized Federated Learning: A Comprehensive Investigation of Strategies and Challenges

Abstract

Personalized Federated Learning (PFL) has emerged as a key paradigm to address statistical heterogeneity in distributed machine learning, tailoring models to individual client data. However, personalization strategies often exacerbate the communication bottleneck, a critical challenge compounded by system heterogeneity across edge devices. This survey provides a comprehensive and structured overview of communication optimization techniques within PFL. This paper proposes a novel taxonomy that categorizes the state-of-the-art into three pillars: (1) architectural methods, such as parameter-efficient learning and structured pruning, that minimize the size of model updates; (2) adaptive process methods, including personalized compression and resource-aware client selection, that optimize the communication protocol; (3) topological frameworks, like clustered and decentralized learning, that restructure collaborative pathways. For each category, this paper analyzes the intricate trade-offs between personalization effectiveness, communication cost, and system complexity. Furthermore, this paper highlights critical challenges related to fairness, robustness, and security that arise from these optimization techniques. Finally, this paper identifies open challenges and chart future research directions, advocating for hybrid, automated frameworks that co-design for efficiency, equity, and security to advance the deployment of robust PFL systems in the real world.