acceptodds
Under review as a conference paper at ICLR 2027

COMPASS: Dual Recalibration for Long-Tailed Adversarial Training

Abstract

Adversarial training (AT) has demonstrated remarkable success on balanced datasets, yet its effectiveness under long-tailed distributions remains challenging, as it necessitates simultaneously improving clean accuracy, overall robustness, and tail-class robustness. Existing methods for long-tailed robustness commonly mitigate class imbalance via prior-based corrections or class-wise perturbation schedules. Nevertheless, they frequently overlook two critical issues: whether a sample is sufficiently reliable in the clean domain to warrant strong adversarial supervision, and whether the final classifier still suffers from bias inherited from the long-tailed training prior. To address these issues, we propose COMPASS, a framework that navigates long-tailed AT through reliable robust supervision and classifier recalibration. COMPASS consists of two modules. Robust Supervision Recalibration (RSR) strengthens adversarial supervision only when class-level robust deficiency is supported by sample-level clean reliability, avoiding excessive perturbations on unstable tail samples. Train-only Classifier Recalibration (TCR) then freezes robust representations and interpolates the original adversarial classifier with a class-balanced recalibrated classifier selected solely from training data. Extensive experiments show that COMPASS substantially improves tail-class robustness and achieves competitive overall performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial training, robustness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_BHRbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes COMPASS for long-tailed adversarial training. It adjusts the per-sample perturbation budget based on clean-prediction reliability and class-level robustness gaps, and then retrains the final classifier on class-balanced data after training. The paper reports improved tail-class robustness across several datasets and architectures.

Critical strengths

None.

Critical weaknesses

W1. The paper reads more like a technical report than a research paper. The introduction claims that strong perturbations on clean-unstable samples harm long-tailed AT, but the paper never shows that existing methods actually fail for this reason. The only related evidence, Fig. 3(b), is measured within COMPASS after the method is introduced. The paper therefore describes what was done and the resulting numbers, but offers no clear claim or insight that a reader can take away beyond the specific combination of techniques.

W2. The method does not distinguish existing components from the proposed contribution. The prior-calibrated loss is logit adjustment as in BSL, the class-wise perturbation budget and TRADES weight follow CFA, and TCR is essentially cRT with an additional interpolation step. The method section presents all of these as parts of COMPASS without separating them from what is new. As a result, it is unclear what the authors actually propose and what the paper is meant to claim.

W3. It is unclear whether the reliability gate itself helps, although it is the core of the paper’s motivation. The introduction and abstract build the paper around the claim that strong perturbations should be applied only to clean-reliable samples, so the reliability gate is the component that most directly supports this claim. However, in Table 4, comparing “Base + RSR” with “Base + RSR w/o Rel.” shows that the gate mostly increases clean accuracy, while robust accuracy stays about the same. But the gate only makes the training attack budget smaller, and a smaller budget usually gives the same result. Fig. 3(b) also just shows that samples given a larger budget become more robust.

W4. Instead, most of the gain comes from TCR, which trades head robustness for tail robustness. In Table 4, the head-class accuracy implied by the all and tail results drops sharply after TCR, while overall robustness barely improves. The tail gains thus mainly reflect classifier rebalancing rather than more robust representations.

W5. The baselines may not share the same base loss. COMPASS includes a BSL-style loss in Eq. (1), but the other plug-in baselines appear to use standard CE. For example, PGD-AT + RLT closely matches AT + RLT in the RLT paper, not AT-BSL + RLT. If so, part of the gain may come from the BSL loss itself, and AT-BSL + RLT would be a more appropriate baseline.

Minor

  • Tables 1–3 are hard to read, and the overall trend is not visible at a glance. Keeping only the TRADES results in the main text would help.
  • Have the authors considered comparing with PAT (Zhang et al., ICML 2026), a recent long-tailed AT method?
  • No statistical report is provided.
  • In Eq. (5) and Eq. (11), should max be min?
Questions for authors

None

13 comments
AuthorsOpenReview ↗

Reply to Reviewer BHRb– Weakness 1

Response to Reviewer BHRb

We thank the reviewer for the detailed and constructive assessment. The review identifies three issues that were not sufficiently separated in the submission: the empirical basis of the reliability hypothesis, the boundary between inherited techniques and our contributions, and the respective roles of RSR and TCR. We address these points below.

Weakness 1. The paper reads more like a technical report than a research paper. The introduction claims that strong perturbations on clean-unstable samples harm long-tailed AT, but the paper never shows that existing methods actually fail for this reason. The only related evidence, Fig. 3(b), is measured within COMPASS after the method is introduced. The paper therefore describes what was done and the resulting numbers, but offers no clear claim or insight that a reader can take away beyond the specific combination of techniques.

Response to the Weakness 1. We agree that Fig. 3(b) is correlational and, by itself, does not establish the failure mode used to motivate RSR. Our scientific claim is narrower than the original wording: class-level robust deficiency is not sufficient to determine the appropriate adversarial pressure for every sample in that class. The allocation problem has two conditions: where robustness is lacking and which samples can receive stronger supervision. RSR is designed around this distinction.

To test this premise directly, we conducted a controlled intervention on CIFAR10-LT with ResNet-18. The model (RLT+PCC) was retrained from random initialization, and all reliability statistics and group thresholds were derived exclusively from the training set. For a training sample (xi,yi)(x_i,y_i) with logits ziz_i, we compute the clean margin

mi=zi,yi−max⁡k≠yizi,km_i=z_{i,y_i}-\max_{k\neq y_i}z_{i,k}

and its class-relative reliability

\qquad \tau_c=\frac{1}{|\mathcal D_c|}\sum_{i\in\mathcal D_c}m_i,$$ where $\tau_c$ is the mean training margin of class $c$ and $T_r=0.5$. Tail classes are the least frequent 80% of classes. Quintiles of their training reliabilities produce the fixed boundaries 0.157, 0.374, 0.618, and 0.846. We define the low-reliability group as the bottom 40% (reliability at most 0.374) and the high-reliability group as the top 40% (reliability at least 0.618); the middle 20% is excluded to obtain two clearly separated groups. Starting from the same final-epoch model, five paired runs compare a weak perturbation budget of $0.3\epsilon$ with the full budget $\epsilon$, while holding the data, objective, optimizer, and three-epoch adaptation schedule fixed. **(a) Reliability and pre-intervention vulnerability.** Q1--Q5 denote increasing reliability quintiles. Samples is the number of test examples, Mean trust is the group mean reliability, and Attack success is 100 minus PGD-20 accuracy. All accuracy values are percentages. | Bin | Samples | Mean trust | Nat. | PGD-20 | Attack success | |:---:|--------:|-----------:|-----:|-------:|---------------:| | Q1 | 3,122 | 0.042 | 12.46 | 0.10 | 99.90 | | Q2 | 1,378 | 0.261 | 77.14 | 1.09 | 98.91 | | Q3 | 1,245 | 0.497 | 100.00 | 18.31 | 81.69 | | Q4 | 1,110 | 0.735 | 100.00 | 67.48 | 32.52 | | Q5 | 1,145 | 0.937 | 100.00 | 93.45 | 6.55 | **(b) Effect of increasing the perturbation budget.** Weak and Strong use $0.3\epsilon$ and $\epsilon$, respectively. Results are mean +/- standard deviation over five paired runs. Delta is Strong minus Weak in percentage points, and $p$ is from a two-sided paired $t$-test. | Group | Samples | Metric | Weak | Strong | Delta | p | |:------|--------:|:-------|-----:|-------:|------:|---:| | Low | 4,500 | Nat. | 27.52 +/- 0.54 | 26.63 +/- 0.58 | -0.89 | 0.042 | | Low | 4,500 | PGD-20 | 0.61 +/- 0.11 | 0.50 +/- 0.09 | -0.11 | 0.047 | | High | 2,255 | Nat. | 97.90 +/- 0.55 | 99.22 +/- 0.26 | +1.32 | 0.002 | | High | 2,255 | PGD-20 | 63.16 +/- 1.66 | 68.05 +/- 0.85 | +4.90 | <0.001 | Panel (a) verifies that the training-derived reliability is meaningfully related to adversarial vulnerability: PGD-20 accuracy increases monotonically from 0.10% in Q1 to 93.45% in Q5, while attack success decreases from 99.90% to 6.55%. Panel (b) isolates the effect of perturbation strength. Increasing the budget reduces low-reliability Nat. and PGD-20 accuracy by 0.89 and 0.11 points, respectively, but improves the corresponding high-reliability metrics by 1.32 and 4.90 points. Thus, the same increase in adversarial pressure has opposite effects depending on clean reliability. This result provides the missing empirical premise for reliability-conditioned allocation. We will reorganize the introduction around this finding, define "clean instability" precisely, and present Fig. 3(b) only as a diagnostic visualization rather than causal evidence.
AuthorsOpenReview ↗

Reply to Reviewer BHRb– Weakness 2

Weakness2. The method does not distinguish existing components from the proposed contribution. The prior-calibrated loss is logit adjustment as in BSL, the class-wise perturbation budget and TRADES weight follow CFA, and TCR is essentially cRT with an additional interpolation step. The method section presents all of these as parts of COMPASS without separating them from what is new. As a result, it is unclear what the authors actually propose and what the paper is meant to claim.

Response. Thank you for raising this concern. We respectfully clarify that COMPASS does not claim logit adjustment, class-wise feedback, or classifier retraining as new in isolation. Its contribution lies in two mechanisms that change how robust supervision is allocated during training and how classifier bias is corrected afterward.

RSR innovation and distinction from BSL/CFA. RSR addresses the within-class ambiguity left by BSL/CFA. It factorizes robust allocation into class-level demand and sample-level readiness: class-wise robust deficiency determines where additional supervision is required, while class-relative clean reliability determines through which samples that demand should be realized. Consequently, two examples from the same under-robust class can receive different effective budgets. The novelty is this conditional cross-granularity allocation, rather than class-wise adaptation or confidence estimation alone. The controlled intervention above supports its necessity: the full budget improves high-reliability PGD-20 accuracy by 4.90 points, but decreases both natural and robust accuracy for low-reliability samples.

TCR innovation and distinction from direct classifier retraining. Conventional decoupled retraining freezes the representation, fits a balanced classifier, and deploys it as a replacement for the original classifier. TCR instead treats classifier correction as a robustness-aware selection problem. It preserves the original adversarial classifier and the balanced classifier as two endpoints, interpolates both their weights and biases, and selects the interpolation coefficient solely on a training-derived split. The selection criterion jointly evaluates natural and robust performance over all and tail classes. Thus, the contribution is not interpolation as an arithmetic operation, but a train-only mechanism for determining how much prior correction can be introduced without fully discarding the adversarially learned decision structure.

Unified claim. COMPASS contributes reliability-conditioned robust allocation and robustness-aware partial classifier correction. We will revise the method section, contribution list, abstract, and related work to mark this boundary explicitly and consistently.

AuthorsOpenReview ↗

Reply to Reviewer BHRb– Weakness 3

Weakness 3. It is unclear whether the reliability gate itself helps, although it is the core of the paper’s motivation. The introduction and abstract build the paper around the claim that strong perturbations should be applied only to clean-reliable samples, so the reliability gate is the component that most directly supports this claim. However, in Table 4, comparing “Base + RSR” with “Base + RSR w/o Rel.” shows that the gate mostly increases clean accuracy, while robust accuracy stays about the same. But the gate only makes the training attack budget smaller, and a smaller budget usually gives the same result...

Response. Thank you for raising this point. The reliability gate is designed to prevent class-level robust feedback from assigning an enlarged perturbation budget indiscriminately to all samples of an under-robust class. Its contribution should therefore be assessed by whether it avoids the clean-accuracy cost of uniform class-wise strengthening while retaining its robust benefit.

Table 4 supports this intended role. Relative to RSR without reliability conditioning, the complete RSR improves Nat.(all) and Nat.(tail) by 5.02 and 5.87 points, respectively. Meanwhile, PGD(tail), C&W(tail), and AA(tail) improve by 0.45, 1.97, and 1.19 points. Overall robustness remains comparable: the changes in PGD(all), C&W(all), and AA(all) are -0.67, +0.69, and -0.14 points, respectively. Thus, the reliability gate substantially improves natural accuracy and tail robustness without materially reducing aggregate robustness.

We directly test the reviewer's alternative explanation by replacing the gate with fixed budgets from 0.3ϵ0.3\epsilon to 0.9ϵ0.9\epsilon. We also include two strength-matched controls: Class-mean matched assigns every sample of a class its mean gate value, while Reliability shuffled preserves the batch-level distribution of gate values but randomly permutes their sample assignments.

Reliability gate versus non-conditional budget controls on CIFAR10-LT. All entries are accuracies in percent.

AllocationNat.(all)PGD(all)Nat.(tail)PGD(tail)
Uniform 0.3 epsilon70.1027.7966.5021.30
Uniform 0.5 epsilon66.6330.9962.3123.98
Uniform 0.7 epsilon65.7734.5559.3625.35
Uniform 0.9 epsilon64.0736.0857.1626.56
Class-mean matched65.6234.3561.1626.55
Reliability shuffled66.8235.1360.7026.18
Reliability gate (ours)68.0735.3862.2527.52

The fixed-budget sweep exhibits a clear clean--robust trade-off rather than reproducing RSR. Reducing the budget to 0.3 epsilon raises natural accuracy but lowers PGD(all/tail) to 27.79/21.30%; increasing it to 0.9 epsilon recovers PGD(all) but lowers Nat.(all/tail) to 64.07/57.16%. No fixed budget matches the gate's joint result. The two matched controls rule out more specific alternatives: Class-mean matched preserves the average budget of each class, while Reliability shuffled preserves the batch-level budget distribution. Both underperform the gate, indicating that the gain depends on assigning each budget according to the corresponding sample reliability, rather than on the average strength or variability of the training attack alone.

We separately test whether the Base+RSR robust gain can be explained solely by the increase in clean-correct predictions. On the same 10,000 CIFAR10-LT test examples, we compare Base with Base+RSR, using a separately generated 20-step PGD attack for each model. Let CC and RR denote clean-correct and robust-correct predictions, respectively.

Tail-sample state transitions from Base to Base+RSR.

Base stateBase+RSR stateCount
Clean-incorrectClean-correct, PGD-incorrect820
Clean-incorrectClean-correct, PGD-correct86
Clean-correct, PGD-incorrectClean-correct, PGD-correct650
Clean-correct, PGD-correctClean-correct, PGD-incorrect217
Clean-correct, PGD-correctClean-incorrect6

Among the tail samples misclassified by Base, RSR makes 906 clean-correct: 820 remain adversarially incorrect, while only 86 become both clean- and robust-correct. Meanwhile, RSR converts 650 samples that were already clean-correct but adversarially incorrect into robust-correct predictions. Accounting for the 217 robust-correct samples that become adversarially incorrect and the 6 that become clean-incorrect, the total net increase in robust-correct tail predictions is

(650+86)−(217+6)=513.(650+86)-(217+6)=513.

The net contribution from samples already clean-correct under Base is 650−217−6=427650-217-6=427. Hence, most of the robust improvement comes from correcting adversarial vulnerability in existing clean decisions, rather than from the increase in clean-correct samples. We will revise the paper to describe the gate as reliability-conditioned allocation.

Reviewer_gZQZOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a adversarial training framework named COMPASS, which is designed to simultaneously enhance the overall classification accuracy, adversarial robustness, and robustness of tail categories of the model. This method consists of two core modules: (1) Robust Supervision Recalibration (RSR), which combines the category-level robustness gap and the sample-level clean prediction reliability to dynamically adjust the adversarial perturbation budget and the robust loss weight, avoiding imposing overly strong adversarial supervision on unstable samples; (2) Train-Only Classifier Recalibration (TCR), which re-trains the classifier by freezing the feature extraction network, using balanced sampling by categories to retrain the classifier, and interpolating with the original classifier to alleviate the classification bias introduced by the long-tail distribution. The authors conducted experiments on CIFAR-10-LT, CIFAR-100-LT, and TinyImageNet-LT, combining different network architectures and adversarial training methods. The results show that COMPASS can significantly improve the classification accuracy and adversarial robustness of tail categories, while maintaining competitive overall performance.

Critical strengths
  1. The paper identifies a practical “clean-reliability mismatch” in long-tailed adversarial training.

  2. RSR and TCR are complementary: RSR improves robust representation learning, whereas TCR corrects classifier bias.

  3. The experiments cover multiple datasets, backbones, base learners, attack methods, and ablations, and demonstrate promising gains in tail-class robustness.

  4. Another strength is the flexibility and practical applicability of the proposed framework. COMPASS is designed as a modular approach that can be integrated with different existing adversarial training methods rather than requiring an entirely new training pipeline.

Critical weaknesses
  1. The candidate set A and the data-split ratios used in TCR are not clearly specified.

  2. The authors should clarify the rationale for choosing defaults that are “close to average performance”.

  3. Choosing α\alpha on a training-derived selection set may overestimate TCR gains. This risk may be mitigated by sensitivity analysis across different split ratios.

  4. Incomplete reporting and reproducibility. Computational overhead and code/split details should be added to reduce reproducibility risk.

Questions for authors

Please supplement the manuscript with additional hyperparameter ablation experiments (e.g., the data-split ratios) and add more detailed descriptions of the core components (e.g., the RSR reliability gating and TCR interpolation/selection procedure) to improve clarity and reproducibility.

For other specific questions, please refer to weakness part.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer gZQZ – Weaknesses 1 and 2

We thank the reviewer for recognizing the clean-reliability mismatch addressed by COMPASS and the complementary roles of RSR and TCR. Below, we clarify the complete TCR protocol, explain the rationale behind the default settings, and report additional sensitivity results.

Response to Weakness 1.

The interpolation candidate set is \begin{equation} \mathcal{A}=\{0,0.1,\ldots,1.0\}. \end{equation} For each class, we randomly reserve 10%10\% of its training examples for Dsel\mathcal{D}_{\mathrm{sel}}, subject to an implementation safeguard of at most twenty selected examples. The resulting selection set contains 10--20 examples per class on C10-LT and 5--20 examples per class on C100-LT and T-LT. In total, this rule produces 185 selection and 13,811 recalibration examples on C10-LT, 1,443 and 18,130 on C100-LT, and 2,889 and 36,183 on T-LT. The corresponding global selection fractions are 1.32%1.32\%, 7.37%7.37\%, and 7.39%7.39\%; they are below 10%10\% because the per-class cap is active for frequent classes. The split is generated once with seed 1.

During TCR, the backbone is frozen, and only the final classifier is optimized for five epochs on Drecal\mathcal{D}_{\mathrm{recal}} using class-balanced sampling with replacement. We use SGD with learning rate 0.010.01, momentum 0.90.9, and weight decay 5×10−45\times10^{-4}.

For every candidate α\alpha, we compute class-wise natural and 20-step PGD accuracy on Dsel\mathcal{D}_{\mathrm{sel}}. Their macro averages define Nat.(all), PGD(all), Nat.(tail), and PGD(tail), and we select the coefficient that maximizes the harmonic mean of these four quantities.

We will add these details to the revised paper.

Response to Weakness 2.

Thanks for this question. The phrase ``close to average performance'' was imprecise and we will modify it.

(1) The defaults are design choices that avoid the two degenerate ends of each control. For smin⁡s_{\min} in Eq. (4), a value near zero can nearly remove adversarial supervision from low-reliability samples, whereas smin⁡=1s_{\min}=1 removes reliability gating and assigns them the full class-level budget. We set smin⁡=0.3s_{\min}=0.3 so that every sample retains a nonzero adversarial signal while the effective budget still varies substantially with reliability. Similarly, λ=0\lambda=0 in Eq. (5) removes class-level perturbation adaptation, whereas a large λ\lambda can amplify noisy class-wise robust-accuracy estimates. We therefore use the moderate value λ=0.5\lambda=0.5 and fix the upper scales qϵmax⁡q_{\epsilon}^{\max} and q~βmax⁡\tilde q_{\beta}^{\max} to 2, as reported in the implementation details. The clipping bounds limit the largest class-wise correction and are not tuned separately across datasets. These four settings, smin⁡s_{\min}, λ\lambda, qϵmax⁡q_{\epsilon}^{\max}, and q~βmax⁡\tilde q_{\beta}^{\max}, remain unchanged across datasets, architectures, and base learners.

(2) The sensitivity plots evaluate stability around these fixed defaults; they are not used to select the best point separately for each setting. That is to say, our intention was not to report the optimal point selected through hyperparameter search, but rather to convey that the method is relatively stable with respect to the choice of hyperparameters.

AuthorsOpenReview ↗

Reply to Reviewer gZQZ – Weaknesses 3 and 4

Response to Weakness 3

We thank the reviewer for this suggestion. We evaluated TCR under five training-derived selection splits, with each result reported as the mean and standard deviation over three independent runs. The Main setting uses our default rule: sampling 10% of the examples from each class, with at most 20 selected examples per class. This gives an actual global selection ratio of 1.32% on C10-LT. The other settings use uncapped class-stratified sampling at ratios of 2%, 5%, 10%, and 20%. For each run, α\alpha is selected only on Dsel\mathcal{D}_{\mathrm{sel}}; the official test set is used once to evaluate the selected classifier.

Sensitivity to the training-derived selection split on C10-LT. The three entries under Selected alpha are the coefficients selected in the three independent runs. Accuracy values are percentages reported as mean +/- standard deviation.

SplitSelected alphaNat.(all)PGD(all)Nat.(tail)PGD(tail)
Main0.5, 0.5, 0.667.75 +/- 0.8437.65 +/- 0.5666.01 +/- 0.4636.57 +/- 0.76
2%0.4, 0.4, 0.568.06 +/- 0.6937.65 +/- 0.7565.87 +/- 0.3035.47 +/- 1.10
5%0.2, 0.4, 0.468.65 +/- 0.9337.21 +/- 0.6065.79 +/- 0.5134.86 +/- 1.36
10%0.4, 0.4, 0.568.02 +/- 0.8137.64 +/- 0.7165.80 +/- 0.3835.30 +/- 0.90
20%0.3, 0.4, 0.468.70 +/- 0.1037.33 +/- 0.8165.95 +/- 0.4334.16 +/- 1.46

The resulting performance remains relatively stable across this range of split sizes. Across the five settings, the mean Nat.(all), PGD(all), and Nat.(tail) vary by only 0.95, 0.44, and 0.22 percentage points, respectively, while PGD(tail) remains between 34.16% and 36.57%. Although the selected α\alpha varies across splits and runs, the resulting aggregate metrics vary only modestly. These results alleviate the concern that the reported TCR performance depends on a particular selection ratio. We will add the exact split construction and this sensitivity analysis to the revised manuscript.

Response to Weakness 4

Table 5 in the main paper reports both wall-clock time and peak memory on all three datasets. RSR does not require an additional inner-maximization pass; it calibrates the perturbation budget within the existing attack procedure, with additional operations limited to computing clean margins and periodically updating class statistics. This is why its measured training cost remains close to that of the base learner. TCR is a short classifier-only stage and costs 0.07×0.07\times, 0.06×0.06\times, and 0.05×0.05\times of base training on C10-LT, C100-LT, and T-LT, respectively. The peak memory usage of Full COMPASS also remains comparable to that of the base pipeline.

The submitted code exposes all settings required to reproduce this stage, including the split seed, the per-class selection rule, the number of recalibration epochs, the optimizer, and the interpolation candidates. For further details regarding the split, please refer to our response to Weakness 1.

Thank you for the authors' response. I believe it has addressed my concerns, and I will raise my score.

AuthorsOpenReview ↗

Further Reply to Reviewer gZQZ

Thank you sincerely for your positive feedback and for acknowledging that our response has addressed your concerns. We greatly appreciate your time and consideration.

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

This paper proposes COMPASS for long-tailed adversarial training, aiming to address two challenges: (1) class-level adversarial supervision may overlook sample-level differences within the same class, and (2) adversarially trained models may suffer from classifier bias under long-tailed distributions. The proposed framework consists of Robust Supervision Recalibration (RSR), which combines class-level robust gaps with sample-level clean reliability to adapt perturbation budgets, and Train-only Classifier Recalibration (TCR), which adjusts the classifier after robust representation learning. Experiments on multiple long-tailed benchmarks under various adversarial attacks demonstrate improvements over existing methods.

Critical strengths
  1. Meaningful research problem The paper addresses the joint challenges of class imbalance and robustness imbalance in long-tailed adversarial training. The combination of robust supervision adjustment and classifier recalibration provides a coherent framework.

  2. Comprehensive experimental evaluation The experiments cover multiple datasets, architectures, and adversarial attacks, including AutoAttack. The extensive comparisons and ablation studies provide useful empirical evidence of the framework's effectiveness.

Critical weaknesses

1: The novelty relative to existing methods is insufficiently clarified. Both adaptive perturbation and classifier recalibration have been investigated in prior work. For example, Taming the Long Tail: Rebalancing Adversarial Training via Adaptive Perturbation (CVPR 2026) explores adaptive perturbation for long-tailed adversarial training, while Decoupling Representation Learning and Classifier for Long-Tailed Adversarial Training studies the decoupling of robust representation learning and classifier optimization. The paper should more clearly distinguish RSR from existing adaptive perturbation strategies and explain what methodological novelty TCR introduces beyond previous decoupled training approaches.

  1. The necessity of sample-level reliability is insufficiently demonstrated. The central motivation of RSR is that low-reliability samples may be harmed by excessive adversarial perturbations, whereas high-reliability samples can benefit from stronger supervision. However, the paper does not directly establish this assumption. Although Figure 3(b) shows that reliable samples obtain larger robustness gains from additional supervision, it does not demonstrate that low-reliability samples are more vulnerable to adversarial perturbations or that excessive perturbations cause performance degradation. Further analysis is needed to establish the relationship between sample reliability, adversarial vulnerability, and the proposed supervision adjustment.

  2. The effectiveness and contribution of RSR remain unclear. Figure 3(a) demonstrates that RSR assigns different perturbation budgets according to sample reliability. However, this mainly confirms that the implementation follows the proposed allocation rule, rather than demonstrating that the allocation effectively mitigates harmful adversarial supervision. Moreover, the ablation results suggest that the benefits of sample-level reliability are more pronounced in natural accuracy than in robust accuracy, while TCR contributes substantially to the overall improvement. This raises questions about whether RSR primarily enhances robust representation learning or acts as an additional regularization mechanism.

Questions for authors

Q1. How does RSR differ fundamentally from existing adaptive perturbation methods for long-tailed adversarial training, and what are the specific methodological contributions of TCR beyond prior classifier recalibration approaches?

Q2. Can the authors provide direct evidence that low-reliability samples suffer from excessive adversarial perturbations under conventional training and that RSR alleviates this problem?

Q3. Can the authors further isolate the contributions of RSR and TCR, particularly in terms of robust accuracy? Does the reliability-guided mechanism provide substantial benefits beyond natural accuracy improvements?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer fSeA – Q1. The novelty relative to existing methods is insufficiently clarified.

Thank you for this concern. COMPASS goes beyond adaptive perturbation or classifier retraining considered separately: it introduces two decisions that the cited approaches do not resolve. At the representation-learning stage, existing class-aware perturbation methods determine how strongly a class should be attacked, but not which samples within that class can safely support the additional adversarial pressure. At the classifier stage, conventional decoupled retraining removes class-prior bias by replacing the original classifier, but does not determine how much debiasing can be introduced without discarding adversarially learned decision structure. RSR and TCR are designed specifically to address these two unresolved decisions. We will make this distinction explicit throughout the revised paper.

The contribution of RSR is a conditional factorization of robust supervision into class-level need and sample-level trust. Methods based on class frequency or class-wise robust accuracy determine whether a class should receive stronger supervision, but apply that decision uniformly within the class. RSR instead uses \begin{equation} \epsilon_i^{(t)} = \epsilon q_{y_i,\epsilon}^{(t)} \left[s_{\min}+(1-s_{\min})r_i^{(t)}\right], \end{equation} where qyi,ϵ(t)q_{y_i,\epsilon}^{(t)} measures the robust deficiency of class yiy_i, while ri(t)r_i^{(t)} measures whether sample ii has a sufficiently stable clean decision to support the resulting perturbation. Consequently, two samples from the same under-robust class can receive different effective budgets. RSR therefore changes the granularity and condition of robust allocation: class feedback determines where robustness is lacking, and clean reliability determines through which samples that deficiency should be addressed.

TCR differs from existing classifier retraining methods in three important aspects. First, existing methods typically freeze the representation, train a new classifier with class-balanced data, and completely replace the original classifier. TCR instead retains the original adversarially trained classifier and uses the balanced classifier only for partial correction. Second, complete replacement focuses mainly on removing class-frequency bias, but may discard decision boundaries that contribute to adversarial robustness. TCR controls this trade-off by interpolating the original and balanced classifiers: \begin{equation} W_{\alpha}=(1-\alpha)W_0+\alpha W_{\mathrm{bal}}, \qquad b_{\alpha}=(1-\alpha)b_0+\alpha b_{\mathrm{bal}}. \end{equation} Third, existing replacement-based methods directly deploy the retrained classifier, whereas TCR determines the correction strength α\alpha on a training-derived selection split by jointly evaluating clean and robust performance over both the full class set and the tail classes. The contribution of TCR is therefore not classifier retraining itself, but a train-only and robustness-aware mechanism that determines how strongly the final classifier should be recalibrated while preserving useful robust decision structure. We will revise the manuscript so that these two precise contributions are consistently separated from the inherited components.

AuthorsOpenReview ↗

Reply to Reviewer fSeA – Q2. The necessity of sample-level reliability is insufficiently demonstrated.

We agree that Figure 3(b) shows an association but does not isolate the effect of perturbation strength. We therefore add a controlled intervention on CIFAR10-LT with ResNet-18. A model is retrained from random initialization; reliability thresholds are derived only from training data. Starting from the same final-epoch model, five paired runs compare budgets of 0.3ϵ0.3\epsilon and ϵ\epsilon for low- and high-reliability tail samples while holding the data, objective, optimizer, and three-epoch adaptation schedule fixed.

All grouping rules are fixed before test evaluation. For each training sample, we compute the clean margin mi=zi,yi−max⁡k≠yizi,km_i=z_{i,y_i}-\max_{k\neq y_i}z_{i,k} and class-relative reliability

\qquad \tau_c=\frac{1}{|\mathcal D_c|}\sum_{i\in\mathcal D_c}m_i,$$ where $\tau_c$ is the mean training margin of class $c$ and $T_r=0.5$. Tail classes are the least frequent 80% of classes. Quintiles of their training reliabilities give the fixed internal boundaries $\\{0.157,0.374,0.618,0.846\\}$. For intervention, low reliability is $r_i\leq0.374$ (bottom 40%), high reliability is $r_i\geq0.618$ (top 40%), and the middle 20% is excluded. These thresholds are applied unchanged to test samples. Before intervention, PGD-20 accuracy rises monotonically from 0.10% in Q1 to 93.45% in Q5, while attack success falls from 99.90% to 6.55%, confirming that reliability tracks adversarial vulnerability. **(a) Reliability and pre-intervention vulnerability.** Samples is the number of test images, Mean trust is the group mean reliability, and Attack success is 100 minus PGD-20 accuracy. All accuracy values are percentages. | Bin | Samples | Mean trust | Nat. | PGD-20 | Attack success | |:---:|--------:|-----------:|-----:|-------:|---------------:| | Q1 | 3,122 | 0.042 | 12.46 | 0.10 | 99.90 | | Q2 | 1,378 | 0.261 | 77.14 | 1.09 | 98.91 | | Q3 | 1,245 | 0.497 | 100.00 | 18.31 | 81.69 | | Q4 | 1,110 | 0.735 | 100.00 | 67.48 | 32.52 | | Q5 | 1,145 | 0.937 | 100.00 | 93.45 | 6.55 | **(b) Effect of increasing the perturbation budget.** Weak and Strong use $0.3\epsilon$ and $\epsilon$, respectively. Results are mean +/- standard deviation over five paired runs. Delta is Strong minus Weak in percentage points, and $p$ is from a two-sided paired $t$-test. | Group | Samples | Metric | Weak | Strong | Delta | p | |:------|--------:|:-------|-----:|-------:|------:|---:| | Low | 4,500 | Nat. | 27.52 +/- 0.54 | 26.63 +/- 0.58 | -0.89 | 0.042 | | Low | 4,500 | PGD-20 | 0.61 +/- 0.11 | 0.50 +/- 0.09 | -0.11 | 0.047 | | High | 2,255 | Nat. | 97.90 +/- 0.55 | 99.22 +/- 0.26 | +1.32 | 0.002 | | High | 2,255 | PGD-20 | 63.16 +/- 1.66 | 68.05 +/- 0.85 | +4.90 | <0.001 | Increasing the budget reduces low-reliability Nat. and PGD-20 by 0.89 ($p=0.042$) and 0.11 points ($p=0.047$), respectively, but improves their high-reliability counterparts by 1.32 ($p=0.002$) and 4.90 points ($p<0.001$). Thus, the same increase in adversarial strength benefits stable predictions but significantly degrades both natural and robust accuracy for unreliable tail samples. This opposing response directly supports reliability-conditioned RSR over uniform perturbation strengthening.
AuthorsOpenReview ↗

Reply to Reviewer fSeA – Q3. The effectiveness and contribution of RSR remain unclear.

We agree that visualizing the assigned budgets verifies the implementation but not their effect. We therefore separate two questions: whether RSR improves the adversarially trained model before TCR, and whether its robust gain merely follows from higher clean accuracy. Table 4 addresses the first question. Before any classifier recalibration, adding RSR to the base improves PGD(all)/PGD(tail) from 32.13/22.99% to 35.38/27.52%; the gains persist under C&W (32.68/22.27% to 34.43/26.26%) and AutoAttack (30.71/20.32% to 31.89/23.73%). Holding TCR fixed yields the same conclusion: Full COMPASS improves over Base+TCR by 1.12/1.51, 1.04/2.35, and 0.96/1.37 percentage points under PGD, C&W, and AutoAttack (all/tail), respectively. Thus, TCR contributes strongly, but it does not account for the pre-TCR robust improvement.

To address the second question directly, we add a paired state-transition analysis on all 10,000 CIFAR10-LT test examples. Model A is the base model, and Model B is Base+RSR. Each model is evaluated using its own 20-step PGD attack with the same random initialization. Let CC and RR denote correct clean and adversarial predictions. The three mutually exclusive states are W (¬C\neg C; wrong clean prediction), C-W (C∧¬RC\wedge\neg R; correct clean but wrong adversarial prediction), and C-C (C∧RC\wedge R; correct clean and adversarial predictions).

Performance in the paired analysis. Nat. and PGD are clean and PGD-20 accuracy. Conditional PGD is PGD-20 accuracy evaluated among clean-correct samples. Delta is B minus A in percentage points. All values are percentages.

SubsetModelNat.PGD-20Conditional PGD
AllA: Base63.5231.4849.56
AllB: Base+RSR67.8733.9650.04
AllDelta+4.35+2.48+0.48
TailA: Base55.5919.6335.30
TailB: Base+RSR62.1426.0441.90
TailDelta+6.55+6.41+6.60

State transitions from Model A to Model B. NsAN_s^A is the number of samples in source state ss under A. Each transition cell reports the sample count followed by its percentage within the corresponding source state.

SubsetA stateN_s^AB: WB: C-WB: C-C
AllW3,6482,730 (74.84%)832 (22.81%)86 (2.36%)
AllC-W3,204460 (14.36%)2,073 (64.70%)671 (20.94%)
AllC-C3,14823 (0.73%)486 (15.44%)2,639 (83.83%)
TailW3,5532,647 (74.50%)820 (23.08%)86 (2.42%)
TailC-W2,877376 (13.07%)1,851 (64.34%)650 (22.59%)
TailC-C1,5706 (0.38%)217 (13.82%)1,347 (85.80%)

We focus on the tail rows to separate robustness improvements from changes in clean classification. Under Model A, 2,877 tail samples are clean-correct but PGD-incorrect (C-W). Model B converts 650 of them into clean- and PGD-correct predictions (C-C), so these gains occur on samples that Model A already classified correctly in the clean domain. Model B also converts 86 samples from clean-incorrect (W) directly to C-C. Conversely, among samples that are C-C under Model A, 217 lose adversarial correctness and become C-W, while 6 lose clean correctness and become W. Model B therefore gains 650+86 robust-correct predictions and loses 217+6, giving the following net increase:

(650+86)−(217+6)=513.(650+86)-(217+6)=513.

Of these 513 additional robust-correct predictions, 650−217−6=427650-217-6=427 (83.2%) arise from the net correction of samples that were already clean-correct under A, whereas only 86 arise from samples that were clean-wrong under A. Consistently, tail P(R | C) increases by 6.60 points, from 35.30% to 41.90%; the overall conditional rate also increases from 49.56% to 50.04%. Hence, the robust gain cannot be explained solely by the increase in clean accuracy: most of the tail gain comes from converting previously vulnerable, clean-correct samples into robust-correct predictions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.