Ensuring fairness of machine learning (ML) algorithms is becoming an increasingly important mission for ML service providers. This is even more critical and challenging in the federated learning (FL) scenario, given a large number of diverse participating clients. Simply mandating equality across clients could lead to many undesirable consequences, potentially discouraging high-performing clients and resulting in sub-optimal overall performance. In order to achieve better equity rather than equality, in this work, we introduce and study proportional fairness (PF) in FL, which has a deep connection with game theory. By viewing FL from a cooperative game perspective, where the players (clients) collaboratively learn a good model, we formulate PF as Nash bargaining solutions. Based on this concept, we propose PropFair, a novel and easy-to-implement algorithm for effectively finding PF solutions, and we prove its convergence properties. We illustrate through experiments that PropFair consistently improves the worst-case and the overall performances simultaneously over state-of-the-art fair FL algorithms for a wide array of vision and language datasets, thus achieving better equity.

本文介绍并研究了联邦学习中的公平性问题，提出了一种新的公平性概念——比例公平性，并通过 Bargaining Games 建立联系，进而提出了一种简单易实现的算法 PropFair，用于在联邦学习中寻找比例公平的解，该算法可以在视觉和语言数据集上进行广泛的实验证明，平衡了所有客户端和最差的 10% 客户端的平均性能。

联邦学习中的比例公平性