Client-level fair federated learning has gained attention as a way to balance utility across heterogeneous participants. Approaches such as q-fair federated learning (q-FFL), agnostic federated learning, and tilted empirical risk minimization aim to prevent the model from favoring well-represented clients at the expense of others.

However, these fairness mechanisms may introduce new vulnerabilities. A malicious or faulty client—often called a Byzantine client—could exploit the reweighting or aggregation rules to push the model in a harmful direction while still appearing to contribute fairly.

The new work proposes FUSE, a fairness-aware update screening method that filters client updates before aggregation. By integrating fairness considerations into the screening process, FUSE aims to block Byzantine exploitation without sacrificing the original fairness objectives. The abstract does not provide experimental details, so the reported effectiveness and trade-offs remain to be examined in the full paper.