acceptodds
Under review as a conference paper at ICLR 2027

FedHRA: Federated Hard-Reliable Recovery and Fair Aggregation for Person Re-Identification

Abstract

Federated learning (FL) enables privacy-preserving person re-identification (ReID) without sharing raw pedestrian images. However, severe ReID domain skew causes client-wise performance unfairness: global models often favor easy clients with clear pedestrian appearances, while hard clients affected by occlusion, blur, poor illumination, and complex backgrounds remain under-optimized. Exist- ing fair FL methods typically compensate under-performing clients by increas- ing their aggregation weights or suppressing conflicting updates. Such strategies are unreliable for ReID because hard-client updates contain both useful domain- specific identity cues and noisy visual components. Blindly amplifying them may strengthen noise, whereas aggressively removing conflicts may discard valuable hard-domain information. Fair federated ReID therefore requires reliability-aware protection of hard-client updates rather than uniform compensation. We propose FedHRA, a Federated Hard-Reliable Recovery and Fair Aggregation framework. FedHRA identifies reliable hard clients by jointly measuring client hardness and update reliability, aligns their updates with the easy-client consensus, and selec- tively recovers reliable hard-domain components. Reliability-guided aggregation further prevents the corrected updates from being diluted. Experiments on four ReID benchmarks with ResNet50 and ViT demonstrate that FedHRA improves hard-client retrieval, overall performance, and cross-client fairness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fair Federated, Person Re-Identification

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_vpWUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses client-wise performance unfairness in federated person re-identification (ReID) under domain heterogeneity. The authors argue that conventional aggregation methods may suppress useful hard-domain information or amplify unreliable updates, leading to performance disparities. To address this, they propose FedHRA, which integrates reliability-aware client scoring, consensus-guided update projection and recovery, and adaptive aggregation. Experiments on four ReID datasets with ResNet50 and ViT backbones report notable improvements in average retrieval performance and worst-client accuracy.

Critical strengths
  1. Meaningful Problem and Reasonable Motivation. The paper addresses an important client-wise fairness issue in federated ReID, with a reasonable and well-motivated problem formulation.

  2. Promising Experimental Results. The experiments cover multiple ReID datasets and two backbone architectures (ResNet50 and ViT), demonstrating notable improvements in retrieval performance, particularly for underperforming clients.

Critical weaknesses
  1. The central motivation and mechanism of FedHRA remain insufficiently validated. The paper claims that the dilution of reliable hard-domain updates and unreliable fairness compensation are key causes of client-wise performance disparity. However, no direct empirical evidence is provided to establish these causes. More importantly, although FedHRA claims to selectively recover reliable identity-discriminative information from hard-client updates, its recovery mechanism simply scales the entire residual using a client-level reliability score, without distinguishing useful identity cues from noisy components. The current ablations only demonstrate performance improvements and do not establish that the proposed reliability score identifies meaningful updates or that the recovered residuals contain useful hard-domain information rather than noise. Therefore, I think the experimental evidence does not adequately substantiate the paper's central motivation and claimed mechanism.
  2. Inconsistent results and insufficient experimental details. Figure 1 and Table 1 report inconsistent results for FedFVand FedHRA Moreover, Figure 1 labels a baseline as FedAvg, although its reported values match FedPav in Table 1. Additionally, many essential implementation details are missing, including the optimizer, learning rate, batch size, local training epochs, key hyperparameter settings.
Questions for authors
  1. Could the authors provide more direct experimental evidence to validate the central claims and proposed mechanisms?

  2. Could the authors clarify the discrepancies between Figure 1 and Table 1? Are they due to different experimental settings or a misunderstanding on my part?

Reviewer_e4tWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates performance fairness issue in federated person re-identification caused by domain skew. The proposed method first identifies hard but reliable clients, whose updates are aligned with the easy-clients. Also, the method prevents corrected hard client updates by reliability-guided global aggregation. Experimental results tested on four benchmark datasets show that the proposed method improves the average performance and client-wise fairness compared with the existing methods.

Critical strengths

The attempt to investigate fairness in federated person re-identification problem is welcomed.

Critical weaknesses
  • Insufficient discussion about experimental results: The authors propose a method to address fairness issues in federated person re-identification (FedReID). Are the experimental settings in Tables 1 and 2 the same as those commonly used in conventional FedReID? For example, the scores of MEDA in Table 1 are different from those reported in the original paper. If the experimental settings are different, the authors should clearly specify how their evaluation protocol differs from the conventional setting. If the experimental settings are same, existing state-of-the-art FedReID methods that aim to improve the performance without explicitly considering fairness would be expected to achieve better overall performance. However, the reported results do not align with this expectation. The authors should provide a discussion to explain this discrepancy.

  • Redundant writing: Writing could be improved. For example, Section 4.1 is recommended to be merged into the Introduction section to reduce redundancy.

  • Unclear mathematical notations: Mathematical notations are often unclear and/or undefined. For example, Eq (1) should be revised, since the RHS seems to F(w) not min F(w). Also, the definition of local objective f(w) is unclear. Some notations are not defined, e.g., P^pre_j, L^pre_u,k, …

Questions for authors
  • Insufficient discussion about experimental results: The authors propose a method to address fairness issues in federated person re-identification (FedReID). Are the experimental settings in Tables 1 and 2 the same as those commonly used in conventional FedReID? For example, the scores of MEDA in Table 1 are different from those reported in the original paper. If the experimental settings are different, the authors should clearly specify how their evaluation protocol differs from the conventional setting. If the experimental settings are same, existing state-of-the-art FedReID methods that aim to improve the performance without explicitly considering fairness would be expected to achieve better overall performance. However, the reported results do not align with this expectation. The authors should provide a discussion to explain this discrepancy.

  • Writing and presentation issues: Writing could be improved. For example, Section 4.1 is recommended to be merged into the Introduction section to reduce redundancy.

  • Unclear mathematical notations: Mathematical notations are often unclear and/or undefined. For example, Eq (1) should be revised, since the RHS seems to F(w) not min F(w). Also, the definition of local objective f(w) is unclear. Some notations are not defined, e.g., P^pre_j, L^pre_u,k, …

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

This paper studies client-wise performance disparity in federated person re-identification. The authors argue that updates from hard clients may contain both useful identity-related information and noisy components, making simple reweighting or aggressive conflict removal unreliable. To address this issue, the paper proposes FedHRA, which consists of Hard-Reliable Client Scoring, Consensus-Guided Update Projection and Recovery, and Reliability-Guided Fair Aggregation. Experiments are conducted on four ReID datasets with both ResNet50 and ViT backbones. The results show clear improvements in average retrieval performance and on several difficult domains.

Critical strengths
  1. The paper studies an important problem in federated ReID.
  2. The method shows consistent improvements.
Critical weaknesses
  1. The fairness formulation is not sufficiently rigorous. The paper uses raw validation mAP across Market-1501, MSMT17, CUHK03, and CUHK02 to identify under-performing clients, and evaluates fairness using raw cross-domain mAP/R1 disparities. However, these datasets have different intrinsic difficulties. Therefore, lower raw performance does not necessarily mean that a client is unfairly under-optimized by the federated model. The current results more clearly demonstrate hard-domain improvement than performance fairness.

  2. The core Hard-Reliable Client Scoring is not directly ablated. In Table 3, both HRUP and RGA already include the scoring module. Therefore, it is unclear whether S_k = W_k R_k^{\mathrm{loss}} R_k^{\mathrm{val}} is actually necessary, or whether much of the gain comes from simpler performance-based reweighting. This concern is particularly relevant because RGA alone already provides a large improvement.

  3. The residual recovery does not explicitly distinguish useful information from noise. The paper motivates FedHRA by arguing that hard-client updates contain both useful identity cues and noisy components. However, the actual recovery applies a single client-level coefficient pkp_k to the entire residual: \hat{g}_k = g_k^{fv} + p_k \left(g_k - g_k^{fv}\right). Therefore, useful and noisy residual components may be recovered together, and the current evidence does not establish that the recovered residual specifically corresponds to reliable hard-domain information.

  4. The stability of the key mechanism is insufficiently studied. The recovery score is normalized as S_k =\frac{S_k-S_{\min}}{S_{\max}-S_{\min}+\epsilon}. If only one hard-reliable client is selected, i.e.,H_s\right| = 1,] then [S_{\max} = S_{\min}, which leads to S_k = 0, \qquad p_k = 0. Thus, no residual is recovered for that client. Since the experiments contain only four domain-level clients, this case cannot be ruled out. The paper does not report the number of hard/easy clients per round, the distribution of pkp_k, or results over multiple random seeds.

Questions for authors
  1. How do the authors distinguish intrinsic dataset difficulty from unfair under-optimization by the federated model? Can the fairness results be normalized by a dataset-specific reference performance?

  2. Can the authors provide a direct ablation of Hard-Reliable Scoring, for example, comparing WkW_k, WkRklossW_k R_k^{\mathrm{loss}}, WkRkvalW_k R_k^{\mathrm{val}}, and the full WkRklossRkvalW_k R_k^{\mathrm{loss}} R_k^{\mathrm{val}}? This would materially affect the assessment of the necessity of the proposed scoring mechanism.

  3. How many hard-reliable clients are typically selected in each round? Does ∣Hs∣=1\left|\mathcal{H}_s\right| = 1 occur in practice, and if so, how does the resulting pk=0p_k = 0 affect training and final performance?

  4. What evidence shows that the recovered residual mainly contains reliable hard-domain information rather than recovering useful and noisy components together?

Reviewer_2iBvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The author propose FedHRA, a Federated Hard-Reliable Recovery and Fair Aggregation framework for person re-identification. It introduces hard-reliable client scoring, consensus-guided update projection and recovery, and reliability-guided aggregation. Experiments on multiple ReID datasets demonstrate improvements in average retrieval performance and cross-client fairness.

Critical strengths

1、This work focus an important challenge in federated ReID,where heterogeneous client domains may lead to imbalanced results. 2、Promising experimental results, and it achieves consistent improvements over multiple federated learning and fair aggregation baselines. 3、Clear design, each components are reasonably motivated, and the ablation experiments demonstrate their complementary contributions.

Critical weaknesses

1、Limited methodological novelty. The paper does not sufficiently establish a fundamentally new optimization principle beyond existing techniques. 2、More experimental results from diversified aspects should be provided, e.g. scalability, robustness, client sizes, more realistic scenarios. 3、The reliability estimation, the training objectives are not necessarily imply that update contains ID infos, more theoretical analyzation are needed.

Questions for authors

1、How does the proposed reliability score correlate with the actual quality of client updates? 2、Can the authors evaluate FedHRA with more clients, partial participation, and varying domain heterogeneity? 3、Can the authors provide more detailed ablations and multi-seed statistical results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.