FedCUS: A Client-Unperceived Selection Strategy for Data Heterogeneity in Federated Learning

Federated learning (FL) has emerged as a pivotal framework for collaboratively training machine learning models across distributed clients while preserving privacy. However, the statistical heterogeneity inherent in client data distributions poses significant challenges to model convergence and system reliability. While existing client selection strategies partially address concerns about heterogeneity, their reliance on client-reported metrics, such as training loss and utility estimates, introduces vulnerabilities to strategic misreporting, which may compromise model integrity. To overcome these limitations, we present FedCUS, a novel client-unperceived selection strategy that ensures strategic-proof cooperation in heterogeneous FL ecosystems. Departing from conventional approaches, FedCUS implements an opaque selection mechanism by decoupling selection criteria from client visibility, thereby eliminating clients' awareness of the selection criteria and disincentivising dishonest reporting behaviours. The proposed framework leverages a DPP (determinantal point process)-based diversity enhancement module for optimal client selection, utilizes matrix completion to estimate missing entries in the likelihood matrix for DPP and incorporates an age-of-update (AoU) aware temporal weighting mechanism to mitigate the effect of stale local updates. Extensive empirical evaluations demonstrate that FedCUS accelerates convergence compared to existing baselines and maintains robust performance in the presence of untruthful clients.