M2F: Multi-centered Fairness-Aware Federated Learning Framework

Federated learning (FL) is a promising technique to train machine learning models across distributed and privacy-conscious devices. The clients co-train a global model, while their contributions to global model training differ due to inherent heterogeneity in data and capabilities. This induces inequitable incentives for their contributions. Existing incentive mechanisms relying on a single model for incentive allocation often underestimate client contributions when there are significant data discrepancies among them. Therefore, this paper proposes a Multi-centered Fairness-aware FL framework (M2F). It implements a clustering method based on model similarity to construct personalized contribution evaluation adaptively. We also design a multi-dimensional metric to evaluate client quality by considering participation rate, computation ability, and training dataset size. In this design, clients receive a customized variant of the aggregated gradient as an incentive at the end of each training iteration. Experimental results validate that the M2F framework can accurately differentiate clients with heterogeneous datasets and diverse quality by increasing the convergence speed and accuracy gaps among them, hence promoting fairness.