MFMA: A Personalized Federated Learning Method Towards Fairness and Accuracy

Federated learning (FL) is a distributed machine learning paradigm that has gained significant attention for training models on large-scale and heterogeneous data. However, ensuring fairness in FL has emerged as a crucial challenge, particularly in scenarios where the server may exhibit unfair behavior towards clients. Addressing this challenge requires novel approaches in federated optimization. In this paper, we propose a regularization-based personalized federated learning method called MFMA, inspired by the loss amplification mechanism. The primary objective of MFMA is to enhance accuracy while preserving fairness. To achieve this, MFMA applies personalized loss amplification to each client, incorporating it as a regular term in the subsequent round of federated learning. This personalized loss amplification helps guide the training process for devices representing low-representation minority groups without adversely affecting the overall training performance. We conduct extensive experiments on various real and synthetic datasets, comparing the performance of MFMA against state-of-the-art baseline methods. The experimental results demonstrate that MFMA exhibits high competitiveness in terms of both performance and fairness.