Fair-select: a federated learning approach to ensure fairness in selection of participants

Federated Learning (FL) is an emerging paradigm for collaboratively training machine learning models while preserving the privacy of participant datasets. Over recent years, significant contributions have been made in various aspects of FL, including handling data and device heterogeneity, enhancing personalization, reducing communication rounds, and accelerating training. However, the optimal selection of participants during each communication round has been under-explored in existing literature. In this paper, we present a novel approach called Fair-Select, which addresses the critical issue of participant selection to ensure fairness and prevent selection starvation. Our approach begins by estimating the utility of each participant based on their loss values from a fixed set of data samples. The server then continuously monitors improvements in both the utility and influence of individual participants to provide fair opportunities for those that have been previously overlooked. This unique monitoring criterion is the cornerstone of Fair-Select, distinguishing it from existing FL methods. We conduct extensive experiments on publicly available datasets to evaluate the effectiveness of our approach, demonstrating that Fair-Select not only improves overall model performance but also increases the diversity of participating clients in each communication round.