Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cumulative Utility Parity for Fair Federated Learning under Intermittent Client Participation

Stefan Behfar, Richard Mortier

TL;DR

This work tackles fairness in federated learning under intermittent client participation by introducing Cumulative Utility Parity, a fairness notion evaluated over long training horizons rather than per round. It develops availability-aware mechanisms—temporal utility tracking, inverse-availability adaptive sampling, and surrogate updates—to ensure that clients receive comparable long-term benefit per participation opportunity, with formal guarantees (Lemmas and Theorems) on convergence and variance reduction. Empirically, using temporally skewed non-IID CIFAR-10 benchmarks and real device availability traces, the approach improves representation parity and reduces dispersion in cumulative benefit while maintaining near-optimal predictive performance. The findings highlight the importance of temporal fairness considerations in FL and offer practical, provable methods for equitable participation across diverse clients.

Abstract

In real-world federated learning (FL) systems, client participation is intermittent, heterogeneous, and often correlated with data characteristics or resource constraints. Existing fairness approaches in FL primarily focus on equalizing loss or accuracy conditional on participation, implicitly assuming that clients have comparable opportunities to contribute over time. However, when participation itself is uneven, these objectives can lead to systematic under-representation of intermittently available clients, even if per-round performance appears fair. We propose cumulative utility parity, a fairness principle that evaluates whether clients receive comparable long-term benefit per participation opportunity, rather than per training round. To operationalize this notion, we introduce availability-normalized cumulative utility, which disentangles unavoidable physical constraints from avoidable algorithmic bias arising from scheduling and aggregation. Experiments on temporally skewed, non-IID federated benchmarks demonstrate that our approach substantially improves long-term representation parity, while maintaining near-perfect performance.

Cumulative Utility Parity for Fair Federated Learning under Intermittent Client Participation

TL;DR

This work tackles fairness in federated learning under intermittent client participation by introducing Cumulative Utility Parity, a fairness notion evaluated over long training horizons rather than per round. It develops availability-aware mechanisms—temporal utility tracking, inverse-availability adaptive sampling, and surrogate updates—to ensure that clients receive comparable long-term benefit per participation opportunity, with formal guarantees (Lemmas and Theorems) on convergence and variance reduction. Empirically, using temporally skewed non-IID CIFAR-10 benchmarks and real device availability traces, the approach improves representation parity and reduces dispersion in cumulative benefit while maintaining near-optimal predictive performance. The findings highlight the importance of temporal fairness considerations in FL and offer practical, provable methods for equitable participation across diverse clients.

Abstract

In real-world federated learning (FL) systems, client participation is intermittent, heterogeneous, and often correlated with data characteristics or resource constraints. Existing fairness approaches in FL primarily focus on equalizing loss or accuracy conditional on participation, implicitly assuming that clients have comparable opportunities to contribute over time. However, when participation itself is uneven, these objectives can lead to systematic under-representation of intermittently available clients, even if per-round performance appears fair. We propose cumulative utility parity, a fairness principle that evaluates whether clients receive comparable long-term benefit per participation opportunity, rather than per training round. To operationalize this notion, we introduce availability-normalized cumulative utility, which disentangles unavoidable physical constraints from avoidable algorithmic bias arising from scheduling and aggregation. Experiments on temporally skewed, non-IID federated benchmarks demonstrate that our approach substantially improves long-term representation parity, while maintaining near-perfect performance.
Paper Structure (21 sections, 89 equations, 3 figures, 2 tables, 2 algorithms)

This paper contains 21 sections, 89 equations, 3 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Model Design
  3. Proof.
  4. Proof.
  5. Empirical Evaluation
  6. Fairness via Temporal Utility Tracking
  7. Adaptive Sampling and Participation Balance
  8. Impact of Surrogate Updates
  9. Baseline comparison
  10. Metrics for evaluation
  11. Baseline evaluations
  12. Conclusion
  13. Quantitative bound on normalized-utility deviation under inverse-availability sampling
  14. Variance reduction under inverse-availability compensation
  15. Assumptions.
  16. ...and 6 more sections

Figures (3)

  • Figure 1: Distribution of device availability percentages for the trace data yang2021characterizing, where device availability percentage is defined as the percentage of time between the first and last times a device was seen to be live and available to perform FL, i.e., was charging and connected to Wi-Fi. Out of 1000 devices in the trace, 213 were available for $<$5% of the time, and over 60% were available for less than half the time.
  • Figure 2: Accuracy and fairness variance versus round number using random sampling.
  • Figure 3: Accuracy and fairness variance versus round number using inverse-availability × missed-round reweighting ($\lambda=0.7$).