Mitigating Learner-Client Distribution Mismatch in Federated Learning

Federated learning (FL) often assumes that the aggregated data from diverse clients accurately represents the central learner's target distribution, or that all clients contribute positively. This research challenges that assumption, highlighting a significant "learner-client distribution mismatch" that can hinder FL performance. Traditional client selection methods often overlook this misalignment, and existing transfer learning approaches typically require direct data access or one-shot aggregation. The paper introduces a dynamic, influence-aware client selection framework designed to mitigate these impacts. By maintaining a small proxy dataset at the learner, the framework estimates each client's potential utility to the learner's optimization objective using proxy influence signals. This allows for prioritizing the most informative clients while controlling for statistical noise and data heterogeneity. Experiments demonstrate that this approach consistently outperforms static and dynamic baselines, leading to faster convergence and higher accuracy in heterogeneous data environments.