FairMoE-FL: A Communication-Efficient and Fair Federated Mixture-of-Experts Framework

Document Type

Conference Proceeding

Source of Publication

IEEE International Conference on Communications

Publication Date

5-24-2026

Abstract

Federated Learning (FL) enables collaborative model training across distributed clients without sharing raw data but suffers from communication overhead and unfair performance due to client heterogeneity. We propose FairMoE-FL, a novel FL framework that integrates a Mixture-of-Experts (MoE) architecture with fairness-aware optimization. FairMoE-FL employs sparse expert activation and dual-level routing with lightweight local adapters, reducing communication while preserving personalization. To ensure equitable outcomes, we introduce fairness-regularized objectives that penalize accuracy disparity, balance expert usage, and incorporate communication cost constraints. Extensive experiments on MNIST and CIFAR-10 show that FairMoE-FL consistently improves accuracy and fairness across clients while transmitting fewer parameters than baseline methods. These results highlight the potential of conditional computation with fairness-aware optimization for scalable, efficient, and equitable FL in heterogeneous environments.

ISBN

[9798319542090]

ISSN

1550-3607

Publisher

IEEE

Disciplines

Computer Sciences

Keywords

Communication Efficiency, Fairness, Federated Learning, Mixture-of-Experts

Scopus ID

105045374707

Indexed in Scopus

yes

Open Access

no

Share

COinS