Cumulative Utility Parity for Fair Federated Learning under Intermittent Client Participation

Intermittent participation creates hidden fairness problems: The paper reveals that existing fairness measures in federated learning can appear fair on a per-round basis while systematically disadvantaging clients who participate less frequently—a problem that goes undetected by current fairness metrics.

Fairness should measure long-term benefit, not just immediate performance: Rather than equalizing accuracy across training rounds, the researchers propose evaluating whether clients receive comparable long-term benefits per opportunity to participate, accounting for the fact that some clients are simply unavailable more often than others.

Physical constraints versus algorithmic bias matter: The paper distinguishes between unavoidable limitations (like a device being offline) and avoidable algorithmic choices (like how the system schedules or combines client contributions), enabling more targeted fairness improvements that don't penalize clients for circumstances beyond their control.

The proposed approach works in practice: Experiments on real federated learning scenarios with uneven participation patterns show that the new fairness principle substantially improves representation equity for intermittently available clients while maintaining near-perfect overall system performance.

Real-world federated systems have structural fairness challenges: The paper highlights that participation patterns in actual federated learning deployments are often correlated with data characteristics or resource constraints, meaning fairness issues aren't random but systematically affect certain groups of users.