FADE: Federated Aggregation with Discrimination Elimination

In this work, we investigate how unfair updates with opposing biases can cancel each other out during aggregation in federated learning (FL), leading to a fairer overall model from a group fairness perspective. We analytically and empirically analyze this Federated Aggregation with Discrimination Elimination (FADE) phenomenon, considering both linear and nonlinear models. In addition, we build on this observation and introduce two novel fairness-aware FL aggregation strategies. The first strategy, FADE-OptW, uses sequential optimization to optimize weights assigned to each client based on their fairness levels. The second approach, FADE-SSP, identifies the optimal subset of clients that minimizes the weighted average fairness level at each round along the convergence path, and for a given metric. Our experiments demonstrate significant improvements in fairness, achieving up to a 60% reduction in discrimination compared to standard FedAvg-based FL. We achieve these gains while maintaining the model's predictive performance on highly heterogeneous client data distributions.