acceptodds
Withdrawn from ICLR 2027

Reward-Free Class-wise Fairness with Global Performance Preservation in Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Collaborative fairness in federated learning concerns how participants’ model benefits relate to their contributions. Under label skew, a single client score can obscure substantial differences in class participation. We study participation-aware class utility allocation together with its effect on shared training. Our framework, SICAR, constructs class directions from relative client distributions on the probability simplex, without separately valuing each class and mapping those values to rewards. This reward-free direction construction is followed by a response sensitive classifier fit and an explicit rank-based amplitude rule. We then subtract each saved additive perturbation from the trained client model before aggregation. The cancellation removes the injected offset, while our analysis characterizes the remaining effect of local optimization. This allows clients to receive differentiated class-wise utility while substantially decoupling the fairness intervention from global model optimization. Experiments show that the proposed framework maintains stable contribution–utility alignment while recovering global model performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Collaborative Fairness, Class-Wise Model Rewards, Probability Simplex, Non-IID Data

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dhvaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses a practical problem in federated learning: allocating collaborative performance gains based on actual contributions, which could have significant implications for encouraging edge devices to participate in FL. The authors implement the SICAR algorithm in three steps: First, generate reward directions based on the class distribution; second, map class targets to classifier parameter perturbations; and third, perform known-offset cancellation after training. The approach appears generally sound, and the experimental results validate its effectiveness to a certain extent.

Critical strengths
  1. The issue addressed in this paper may serve as an important foundation for promoting distributed cooperative learning.
  2. The article has a clear logical structure, and the proposal is well-defined.
Critical weaknesses
  1. How can the authors demonstrate that relative proportions are more appropriate than absolute numbers for characterizing a client’s contribution to a particular label? The contribution of labels is clearly strongly correlated with both the quantity and quality of data; I believe it is unreasonable to define it based solely on data quality or quantity, unless the author can provide relevant theoretical support or sufficiently persuasive experimental evidence.
  2. Fairness is achieved by reducing the performance of all clients, which raises serious practical concerns. The goal of cooperation is to achieve higher accuracy and better generalization performance, but it is clear that the cancellation operation significantly interferes with this goal, as confirmed by the experimental results. This raises the question: Does the improvement in fairness stem primarily from selective degradation, rather than from providing better models to high-contribution clients?
  3. The authors don't conduct a systematic study of the cooperation among FL clients of different scales, including but not limited to data scale and user count; and don't conduct a systematic study of data that is non-IID to varying degrees, as well.
Questions for authors

please refer to "weaknesses".

ps:

  1. Just for discussion: Why can't we just use class-wise logit bias adjustment? This kind of method is simpler and more efficient, and there is no experimental evidence to suggest that it is less effective than the approach proposed in this paper.
Reviewer_XAA2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SICAR, a framework for class-wise model allocation in federated learning. It constructs client-specific class directions using centered log-ratio contrasts. These directions are then mapped to classifier perturbations through a regularized response fit in a subspace orthogonal to retained clean-update directions. Clients train from the perturbed models, and the server subtracts the saved perturbations before aggregation. The experiments evaluate stable contribution–utility alignment and global model performance.

Critical strengths

1.This paper provides a coherent framework for studying how differentiated model allocation interacts with shared training.

2.The paper successfully designs a series of experiments across classic datasets (e.g. the CIFAR-100 and Tiny-ImageNet) which support the proposed mechanism.

3.The authors explicitly acknowledge the remaining optimization effects, reduced dispatched utility, weak SVHN results, and limitations of the baseline protocols. This transparency helps delimit the claims supported by the experiments.

Critical weaknesses

1.It is still unclear for me why relative class proportions are an appropriate basis for fair rewards. The method decides which classes to favor mainly from each client’s label distribution. However, a larger proportion of a class does not necessarily mean a greater contribution to that class. The rule also does not directly account for data quality or redundancy. The authors claim that this is a chosen allocation policy, but more evidence is needed to explain why this policy should be considered fair.

2.The main class-alignment metric checks whether the observed changes follow the direction prescribed by the method itself. It shows that the method can implement its target, but does not show that the target produces fair or beneficial outcomes. I think a high correlation is possible even when all classes lose accuracy by different amounts.

3.The experiments cannot fully establish how reliably the method works. For example, subtracting the original perturbation does not completely undo the changes it has caused during local training. In this paper, authors examines this issue through paired training comparisons and local response diagnostics, which are useful. However, the reported response diagnostic only measures selected effects in the classifier and does not fully capture changes in the backbone or their accumulation over training rounds.

Questions for authors

see weakness

Reviewer_r3PBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SICAR, a federated learning mechanism that maps each client's class distribution on the probability simplex to a class-specific classifier perturbation. Class directions are constructed from centered log-ratio geometry between smoothed client distributions and the federation distribution. A labeled server probe estimates class-margin responses, and a ridge fit realizes a scaled target in a response-sensitive subspace orthogonal to retained clean-update directions.

Critical strengths
  1. The paper distinguishes shared-model recovery from dispatched client utility and attempts to connect relative class participation with shared training. This is more informative than reporting only global accuracy.

  2. The algorithm, construction parameters, and complexity discussion are reasonably complete.

  3. The paper analyzes the ridge response fit, spectral shrinkage, a local cancellation criterion, momentum-induced discrepancy, and response-space diagnostics. These analyses are conditional but go beyond purely empirical tuning.

Critical weaknesses
  1. “Reward-free” applies only to class-direction construction. The method is better described as a geometric allocation policy that avoids explicit reward estimation, not as a reward-free contribution-valuation mechanism. The authors acknowledge that class counts are an allocation policy rather than causal valuation, but this weakens the collaborative-fairness and contribution-utility claims.

  2. Table 1 has unequal budgets, and some adapted baselines use historical test-informed scores. Quota is a shared-model surrogate, and the independent FedAvg run differs from paired A on CIFAR-100/ResNet18 without isolating the cause. These make the “best method” reading only descriptive.

  3. Cancellation exactly removes the saved additive offset, but optimization-induced effects remain. The Lipschitz bounds are loose and not estimated. The spectral analysis describes the response subspace but does not provide a verifiable performance guarantee. The theory is useful but insufficient for the stronger claims.

Questions for authors
  1. The independent FedAvg and paired A differ on CIFAR-100/ResNet18. What causes this gap? Can you use paired A as the primary baseline and explain the reference-reuse, initialization, or implementation differences?

  2. Is the cancellation condition in Eq. (18) directly tested?

  3. How do you handle privacy and strategic behavior for class counts and labeled probes?

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

Related papers

Discussion (0)

Sign in to comment.