Towards Generalization Fairness in Federated Learning

Federated learning (FL) has emerged as a new paradigm for privacy-preserving collaborative training in mobile computing, where fairness holds paramount importance. Traditional efforts have largely concentrated on ensuring fairness across clients with different data distributions (i.e., domain- wise generalization fairness). However, the aspect of fairness within classes (class- wise generalization fairness) remains largely unexplored. Thus, an important question arises: <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">is it possible to simultaneously achieve domain- wise and class- wise generalization fairness?</i> Moreover, current approaches often improve model performance on weaker distributions at the cost of compromising performance on stronger ones, introducing another dimension of unfairness. So, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">can we boost performance on weaker distributions without compromising that on stronger ones?</i> To this end, we introduce a global classifier with super logits distillation strategy to achieve comprehensive generalization fairness. The basic idea is to guide local updating by selecting reliable global supervision via a unified global classifier across multiple clients. First, we use these super logits to improve underperforming distributions while preserving the performance of well-performing distributions, promoting domain- wise generalization fairness. Second, we dynamically allocate local distillation loss weights according to the intra-client and inter-client class fairness, accelerating the training of underperforming classes and enhancing class- wise generalization fairness. Comprehensive experiments demonstrate the enhanced fairness and superior performance of our method, highlighting the significance of fairness in mobile computing environments.