A Fair Federated Learning Framework based on Clustering

Abstract

Federated learning is a distributed learning with multiple participants, and the cooperation and competition among participating clients tend to raise the issue of collaborative fairness. If the collaboration among clients is not fair, resulting in resources being monopolized by some advantaged clients, depriving vulnerable clients of participation opportunities, and even causing a “Matthew effect” on client learning resources. To address the above issues, we propose a fair federated learning framework based on clustering. First, we align vulnerable participants into clusters and increase their chances of participating in the federation. The distribution of the global model in clustering is according to the principle of “allocate by work” to balance contributions and profits. We transform the clustering problem into a Liapunov online queue scheduling problem by establishing a bi-objective optimization function. Finally, through theoretical and experimental analysis, the scheme is proved to have good performance in accuracy and fairness.