Assessing the Impacts of Imperfect Datasets on Client Selections in Federated Learning

This paper experimentally studies how quantity skew, label-distribution skew, noisy data, and fairness in client selection affect federated-learning accuracy and convergence. It then proposes a privacy-preserving scoring method for estimating each client’s contribution, aiming to exclude harmful participants without producing an excessively biased selection policy. The supplied abstract says experiments demonstrate the assessment method’s effectiveness, but it provides no datasets, model architectures, baselines, privacy definition, or numerical results, so the strength and generality of the evidence cannot yet be evaluated.