Data-Free Federated Learning Framework Addresses Class Imbalance

A new framework called CELM (class-wise contribution estimation and aggregation via logit maximization) enables data-free client contribution estimation in federated learning (FL). FL allows collaborative training of computer vision models without centralizing data, but practical deployments often suffer from class imbalance and label skew, causing standard aggregation to overfit dominant clients. CELM does not require raw data, client metadata, or auxiliary public datasets. The server probes client updates to obtain class-wise evidence scores, assembles a cross-client evidence matrix quantifying per-class competence and coverage, and computes contribution weights that upweight clients providing strong evidence for underrepresented classes. The resulting aggregation is stable. The paper is available on arXiv with ID 2605.18892.