DFFL: A dual fairness framework for federated learning

Abstract

Federated learning (FL) is an emerging paradigm of distributed machine learning that facilitates collaborative training of a global model across multiple clients while preserving client-side data privacy. However, current equality fairness methodologies aim to maintain a more uniform performance distribution across clients, but they fail to consider the varying contributions of different clients. In contrast, collaboration fairness takes into account the contributions of clients but may exclude low-contributing clients in pursuit of the interests of high-contributing clients. To address these concerns, this paper proposes a novel Dual Fair Federated Learning (DFFL) framework. Specifically, we combine the concept of cosine annealing to evaluate each client’s contribution from two perspectives. Then, we utilize client’s contribution as the aggregation weight of the global model to improve the global model accuracy. Additionally, we introduce a personalized design and utilize client’s contribution as a regularization coefficient to achieve dual fairness. Furthermore, we conduct a theoretical analysis of the convergence of the global model. Finally, through comprehensive experiments on benchmark datasets, we demonstrate that our method achieves competitive predictive accuracy and dual fairness.