Towards High Efficient Federated Learning Systems

Abstract

Federated learning (FL) has shown strong potential in domains such as Intelligent Transportation Systems (ITS), Internet of Things (IoT), and industrial cyber–physical systems (ICPS). However, FL faces three fundamental challenges: communication heterogeneity, statistical heterogeneity, and system heterogeneity. Communication heterogeneity arises from differences in network conditions and transmission capacity among clients, statistical heterogeneity results from non-IID data distributions across clients, and system heterogeneity reflects variations in computational resources and client availability.

Existing studies often address these challenges independently, leaving their interactions insufficiently understood. Among them, communication heterogeneity is particularly critical because it increases communication cost and exacerbates the impact of the other two forms of heterogeneity. This thesis investigates neural network pruning as a central mechanism to mitigate communication costs in FL and develops a progressive research trajectory across increasingly complex environments.

First, pruning is applied at the base system level to reduce communication cost while preserving model accuracy. In large-scale ITS scenarios, a dual pruning mechanism is proposed to adapt to intensified communication demands. To address statistical heterogeneity, a Gaussian Mixture Model (GMM)-based oversampling method is introduced to mitigate slightly skewed label distributions. Finally, a comprehensive pruning framework is developed for complex FL systems that jointly considers communication, statistical, and system heterogeneity.

This work provides both practical solutions for reducing communication cost and a systematic perspective on the interaction of key challenges in FL, supporting scalable deployment in real-world systems.