Unified Group Fairness on Federated Learning

Federated learning (FL) has emerged as an important machine learning paradigm where a global model is trained based on the private data from distributed clients. However, most of existing FL algorithms cannot guarantee the performance fairness towards different groups because of data distribution shift over groups. In this paper, we formulate the problem of unified group fairness on FL, where the groups can be formed by clients (including existing clients and newly added clients) and sensitive attribute(s). To solve this problem, we first propose a general fair federated framework. Then we construct a unified group fairness risk from the view of federated uncertainty set with theoretical analyses to guarantee unified group fairness on FL. We also develop an efficient federated optimization algorithm named Federated Mirror Descent Ascent with Momentum Acceleration (FMDAM) with convergence guarantee. We validate the advantages of the FMDA-M algorithm with various kinds of distribution shift settings in experiments, and the results show that FMDA-M algorithm outperforms the existing fair FL algorithms on unified group fairness. Introduction Federated learning (FL) has emerged as an important machine learning paradigm where distributed clients (e.g., a large number of mobile devices or several organizations) collaboratively train a shared global model while keeping private data on clients (McMahan et al. 2017). However, FL may suffer from fairness problem by disproportionately advantaging or disadvantaging the model performance on different subpopulations, which becomes an increasing concern, especially in some high-stakes scenarios such as loan approvals, healthcare, etc (Kairouz et al. 2019). How to develop a fair FL framework is of paramount importance for both academic research and real applications, and has become an important research theme in recent years (Mohri, Sivek, and Suresh 2019; Li et al. 2019; Wang et al. 2021). A straightforward idea is to apply the methods designed for centralized fairness problem to FL setting. Unfortunately, most of them assume a centralized available training dataset, which infringes the data privacy in FL (Barocas and Selbst 2016; Woodworth et al. 2017; Mehrabi et al. 2021). Recently, some works have been proposed to encourage the federated model to have similar performance over c data of different attributes ai data of newly added clients c′i data of existing clients ci (a) Client-level Fairness (b) Attribute-level Fairness (c) Agnostic Distribution Fairness c1