acceptodds
Withdrawn from ICLR 2027

Cumulative Utility Parity: Fair Learning under Intermittent Client Participation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Intermittent and heterogeneous client availability creates unequal opportunities to contribute to federated learning, a disparity that snapshot metrics for loss and accuracy do not capture and completely miss this unfairness. We introduce cumulative utility parity (CUP), which evaluates each client's accumulated training utility relative to its observed availability. CUP prioritizes clients with larger availability-normalized utility deficits and uses cached model updates as surrogates for unavailable clients. We characterize availability-normalized utility disparity, derive a max–min allocation target, and bound the discrepancy between cached surrogates and fresh updates. Across non-IID CIFAR-10 experiments with two and five labels per client, CUP without caching achieves lower mean temporal utility dispersion than all evaluated baselines in the two-label setting and than FedProx, q-FFL, AFL, and FairFedCS in the five-label setting. The cached CUP variants achieve the lowest mean CV and highest mean Jain index in both settings, together with the highest mean final accuracy in the five-label setting. These findings highlight the potential of availability-aware utility allocation to improve cumulative fairness alongside predictive performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fair Learning, Federated Learning, Availability, Cumulative Utility Parity

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_1mh2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Cumulative Utility Parity (CUP), a fairness-aware federated learning framework for intermittent client participation. It introduces availability-normalized cumulative utility as a fairness objective and prioritizes clients based on historical utility deficits and missed participation opportunities. Cached model updates further compensate for unavailable clients. Theoretical analysis characterizes utility disparity and allocation targets, while experiments on non-IID CIFAR-10 evaluate fairness and predictive performance.

Critical strengths
  1. The paper addresses a meaningful fairness problem by introducing availability-normalized cumulative utility, extending fairness evaluation beyond instantaneous performance.

  2. The experiments use real-world availability traces and demonstrate improved cumulative fairness. The caching mechanism also provides practical benefits with limited runtime overhead.

Critical weaknesses
  1. The methodological novelty is limited, as the proposed scheduling strategy closely resembles existing fairness-aware client selection methods. Its distinction from FairFedCS remains insufficiently justified.

  2. The theoretical analysis does not guarantee that the practical scheduler achieves utility parity. Moreover, the limited experimental settings and lack of cache-enabled baseline comparisons weaken the empirical evidence.

Questions for authors
  1. How does clipped local loss reduction reflect meaningful client contributions under heterogeneous data distributions?

  2. Does the proposed scheduler achieve the theoretical max–min allocation target derived in Lemma 2?

  3. How much of the improvement comes from caching rather than the scheduling mechanism? Would cache-enabled baselines show similar gains?

  4. Do the fairness improvements persist under end-to-end model training and longer communication horizons?

Reviewer_aJgKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes cumulative utility parity (CUP) for federated learning, which measures each client’s accumulated training utility relative to its opportunities to participate. The authors introduce an availability-aware scheduler and a time-decayed cache of absent clients’ updates. The proposed method is evaluated through theoretical analysis and CIFAR-10 experiments.

Critical strengths
  1. CUP explicitly accounts for clients’ participation opportunities, providing a useful fairness formulation for intermittent availability.
  2. The framework clearly separates scheduling, realized-utility accounting, and cached aggregation, without crediting cache reuse as new participation.
  3. The paper provide many theoretical analysis results. The evaluation uses real device availability traces.
Critical weaknesses
  1. The scheduler optimizes clipped reductions in clients’ local training loss, and fairness is then assessed using that same utility measure. This leaves unclear whether CUP improves clients’ held-out performance or service quality, rather than parity on the training signal it directly targets.
  2. The theoretical results clarify the CUP objective, but provide limited guarantees for the proposed algorithm. In particular, the optimal rate allocation assumes fixed mean utilities, while no result establishes that Algorithm 1 realizes these rates under changing utilities and per-round participation constraints. The optimization and cache bounds also retain bias or approximation assumptions whose validity is not demonstrated empirically.
  3. The accuracy curves in Figure 2(a) appear to remain upward-trending at the final round, raising concerns that 50 rounds are insufficient for convergence. Longer runs are needed to determine whether the reported accuracy and fairness advantages persist after sufficient training.
  4. Experiments with only CIFAR-10 is far from sufficient.
Questions for authors
  1. Why is global CIFAR-10 accuracy only around 45–53% with a pretrained ResNet? Please specify the backbone, preprocessing, and BatchNorm handling, and report a centralized classifier-only baseline and longer training runs to clarify the source of this low accuracy.
  2. How does CUP compare with a simple availability-normalized deficit scheduler?
  3. Does caching outperform existing stale-update methods under the same scheduler?
Reviewer_L4fCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a federated learning (FL) framework that aims to improve fairness through cumulative utility parity. The proposed method leverages cached models from temporarily unavailable clients to mitigate fairness disparities caused by heterogeneous client availability. The authors evaluate the effectiveness of their approach on the CIFAR-10 dataset.

Critical strengths
  1. The paper addresses fairness in federated learning, which is an important research problem, particularly in practical FL systems where client availability may vary significantly over time.
  2. The paper attempts to incorporate client availability into fairness-aware federated optimization. This perspective is practically relevant because intermittent client participation is common in real-world FL deployments.
Critical weaknesses
  1. Limited methodological novelty and insufficient justification. The proposed method appears to offer limited methodological novelty, and its technical contributions are not sufficiently distinguished from existing approaches to fairness-aware FL. In its current form, the paper does not convincingly demonstrate the level of innovation and technical depth expected for ICLR. More importantly, the motivation and effectiveness of the proposed cached-model mechanism remain questionable. The authors claim that cached models from unavailable clients can be incorporated to improve fairness. However, it is unclear why reusing potentially outdated models would be preferable to selecting other currently available clients. Furthermore, if a client remains offline for an extended period, its cached model may become increasingly stale, potentially introducing biased or outdated updates that negatively affect global model convergence and performance. Since intermittent client availability is a common characteristic of large-scale FL systems, the authors should provide a more rigorous justification for the necessity and effectiveness of the proposed mechanism.

  2. Insufficient consideration of data heterogeneity and poor presentation quality. The paper overlooks an important challenge arising from the interaction between heterogeneous client availability and non-IID data distributions. In particular, when clients with highly skewed local data distributions exhibit consistently low availability, it remains unclear whether incorporating their cached models can effectively improve fairness without compromising global model performance. This issue is especially important because outdated local updates may introduce additional optimization bias under non-IID settings. Additionally, the experimental evaluation is limited to CIFAR-10, which provides insufficient evidence of the generalizability of the proposed approach across diverse datasets and heterogeneous FL environments.

Questions for authors
  1. The proposed method incorporates cached models from unavailable clients to improve fairness. However, why is reusing cached models preferable to selecting additional currently available clients?
  2. How does the proposed framework handle clients that remain offline for extended periods? In such cases, cached models may become increasingly outdated and potentially introduce biased updates into global aggregation.
  3. How does the proposed method address the interaction between non-IID data distributions and heterogeneous client availability? Specifically, when clients with highly skewed local data distributions exhibit consistently low participation rates, can their cached models still contribute positively to fairness optimization?
  4. The experimental evaluation is limited to CIFAR-10. How well does the proposed method generalize to more challenging datasets and larger-scale FL systems with diverse client availability patterns?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.