acceptodds
Withdrawn from ICLR 2027

AURORA: Attack-Trajectory Uncertainty-Aware Risk Allocation for Long-Tailed Adversarial Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing long-tailed adversarial training methods commonly allocate robust supervision using class-frequency priors or empirical endpoint failures. However, class frequency is not a direct measure of adversarial vulnerability, while endpoint failure rates are unreliable under limited class support and discard how failures develop along the attack trajectory. We propose AURORA, an uncertainty-aware framework for estimating class-wise adversarial risk from endpoint outcomes and intermediate attack predictions. Specifically, uncertainty-Aware Risk Anchoring (URA) applies a Wilson-score upper estimate to endpoint failures, allowing class support to affect allocation through estimation uncertainty rather than a predefined frequency-to-weight mapping. First-Passage Trajectory Refinement (FTR) then measures the normalized first-failure iteration among initially correct samples and uses this path information to refine the endpoint anchor. A bounded correction converts the resulting endpoint-to-path profile into class weights for the native robust objective of existing AT methods. Experiments across multiple long-tailed benchmarks and architectures show that AURORA consistently improves tail robustness while retaining competitive overall performance.

Primary area:
interpretability and explainable AI
Keywords:
long-tailed adversarial training, Uncertainty-Aware Risk Anchoring, First-Passage Trajectory Refinement

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_b6QnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors of this paper propose AURORA, a framework for improving adversarials training on datasets with class imbalance. AURORA consists of two components: uncertainty-aware anchoring (URA) and first-passage trajectory refinement (FTR) that address two limitations of existing adversarial training methods. First, the authors recognise that AT techniques that use failure rates don't account for the fact that a failure rate calculated on a low frequency group is statistically less reliable, so URA uses the Wilson-score upper confidence bound to account for how many observations support the estimated failure rate. Second, FTR uses the earliest attack iteration at which an initially correct prediction turns incorrect, capturing differences in adversarial vulnerability that aren't reflected by final failure rates alone. These two signals are then combined into class-wise weights applied to the underlying AT objective, without introducing an additional loss function. The authors evaluate AURORA on three long-tailed image datasets, across multiple architectures and adversarial training methods, comparing against existing long-tailed adversarial techniques. The authors show AURORA improves tail-class robustness, while generally maintaining competitive overall performance.

Critical strengths
  1. A relatively simple methodology that appears well motivated by limitations of existing techniques. The motivation is interesting, and AURORA seems straightforward to integrate into existing adversarial training procedures.

  2. Fairly extensive evaluations on three datasets, three adversarial training methods, and three different model architectures show AURORA performs well on tail-class metrics, and remains competitive for the overall performance metrics. One small note, the presentation of these results could be improved. Table 2's caption should be consistent with Tables 1 and 3 in explaining the highlighting, and using more distinguishable colours for the two types of highlighting would make the results easier to interpret. The current formatting doesn't quite do the evaluation justice.

Critical weaknesses
  1. There is some uncertainty in my head for whether the iteration at which the adversarial attack achieves misclassification is a true measure of class vulnerability. For example, a particular attack might simply be more effective at misclassifying for certain classes due to differences in the loss landscape or gradient directions, rather than those classes actually being more vulnerable. Although the ablation shows that FTR improves performance, I would have liked to see some study into whether this first-passage behaviour is consistnt across different attack configurations, such as step sizes, number of iterations, or random initialisations. This would help justify the use of attack timing as a measure of vulnerability.

  2. I don't think that the authors justify why the Wilson-score upper bound was chosen over other approaches. Although Table 4 shows that URA improves over using raw failure rates, it would have been useful to compare it against alternatives- for example the Clopper-Pearson confidence bound. This would help to see whether the improvements are specific to Wilson-score.

Questions for authors
  1. Please see W1. It would be useful to see whether this first-passage behaviour is consistnt across different attack configurations, such as step sizes, number of iterations, or random initialisations.

  2. Please see W2. Can the authors comment on why the Wilson score was the chosen upper bound for URA? Would other appraoches such as different confidence bounds produce similar results?

  3. The authors say that their training setup includes prior-aware logit correction and a class-dependent perturbation schedule. Were these same settings used for the competing methods in Tables 1-3? If not, could some of the improvements be due to differences in training setup, not just from AURORA?

Clear answers to the questions above may allow me to reconsider my recommendation more favourably.

Reviewer_jLA7OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper describes class-wise allocation of robust supervision in long-tailed adversarial training (LT-AT). It argues that frequency priors are not a direct measure of adversarial vulnerability and that empirical endpoint failure rates are noisy for small classes and ignore when, along the attack path, a sample fails. The proposed AURORA has two parts. URA replaces each class's training-set endpoint failure rate with the upper end of a Wilson score interval (z=1.96z=1.96), so that classes with fewer samples receive an uncertainty premium. FTR records, for initially correct samples, the first PGD iteration at which the prediction flips, builds a Wilson-upper cumulative first-passage curve per class, and summarizes it by its normalized area DcD_c. The class risk is Rc=Ucendexp⁡(λ ωt dc)R_c = U^{\text{end}}_c \exp(\lambda\,\omega_t\,d_c), where dcd_c is the centered and clipped DcD_c; risks are mean-normalized into next-epoch weights on the robust loss term of PGD-AT, TRADES or AWP. Propositions 1–5 establish reduction to URA at λ=0\lambda=0, bounded refinement, unit-mean normalization, an ordering property and well-posedness. Experiments on CIFAR-10-LT (ratio 50), CIFAR-100-LT and Tiny-ImageNet-LT (ratio 10) with ResNet-18, PreActResNet-18 and WRN-28-10 report clean, PGD-20, C&W and AutoAttack accuracy over all classes and over the 80% least frequent classes, with large tail gains on CIFAR-10-LT (e.g., +9 to +13 points PGD(tail) over the strongest plug-in baselines) and smaller gains elsewhere.

Critical strengths

S1. Cheap, non-invasive plug-in. AURORA reuses predictions the inner maximization already computes, detaches all statistics, and only rescales the native robust term with bounded, unit-mean weights (Props. 2, 3, 11). It needs no auxiliary loss or extra model and adds one hyperparameter (λ\lambda), with an exact fallback to URA at λ=0\lambda=0. If the gains hold under a controlled evaluation, this makes the method easy to adopt across AT objectives.

S2. Broad experimental grid and a controlled estimator ablation. The method is evaluated across three datasets, three architectures and three base learners, with C&W and AutoAttack in addition to PGD. Table 4 is the most informative experiment in the paper: it fixes the training recipe and varies only the class statistic (uniform, raw endpoint failure, CFA-style, URA, AURORA), which is the right design for isolating the contribution of the estimator.

Critical weaknesses

W1. The empirical evidence is internally inconsistent and the main comparison is not controlled. This is the primary reason for my recommendation, because the paper's contribution rests almost entirely on the size of the empirical gains.

(a) Impossible values in the tables. Test sets are balanced and "tail" is the 8 least frequent of 10 classes on CIFAR-10-LT, so head (2-class) accuracy is 5⋅all−4⋅tail5\cdot\text{all} - 4\cdot\text{tail}. For AWP+RLT on C10-LT, Table 1 reports Nat.(all) 57.7% and Nat.(tail) 44.4%, which implies a head clean accuracy of 110.9%. Table 3 (WRN) reports 68.9% / 56.6% for the same column, implying 118.1%. These cannot be correct, which raises the question of how the baseline numbers were produced (re-run, copied from other papers, or mixed).

(b) AutoAttack is weaker than PGD-20 for the PGD-AT baseline in every setting. AutoAttack takes a per-sample worst case over an ensemble that includes APGD-CE (100 iterations), so its robust accuracy should not exceed PGD-20 accuracy beyond noise. Yet for PGD-AT Base, AA > PGD in all nine dataset × architecture combinations, on both "all" and "tail" (e.g., ResNet C10-LT 27.5 vs 26.4; ResNet T-LT 10.9 vs 9.0; WRN T-LT 12.7 vs 10.5). A systematic inversion like this points to different checkpoints, attack settings or transcription errors between columns. In addition, AURORA's own PGD-to-AA drop is larger than most baselines' (C10-LT, all classes: 4.1–5.2 points for TRADES/AWP+AURORA versus roughly 2–3 for most baselines). Loss reweighting is known to induce gradient masking that standard PGD misses: Hitaj et al. (2021) showed GAIRAT's robustness falls from about 55% to 44% under a logit-scaling PGD attack. A method of the same family needs the same scrutiny.

(c) Gains are confounded with the shared training recipe. All AURORA runs use a prior-aware logit correction (τ=1\tau=1) and RLT's class-dependent perturbation schedule (Sec. 4). Table 4's "Uniform Allocation" row (TRADES, ResNet, same recipe, no AURORA weighting) already reaches Nat.(all) 63.7 and PGD(tail) 22.8, which beats every TRADES plug-in baseline in Table 1 on both metrics (best: 62.8 and 21.9, both DAFA). So a substantial part of the headline margin over plug-in methods comes from the recipe, not the estimator, and the paper does not state whether UDR/CFA/DAFA/RLT were given the same recipe. The fair margin attributable to AURORA is the one against uniform allocation (+8.3 PGD(tail) in Table 4), and this is reported for a single setting without seeds or standard deviations. λ=0.8\lambda=0.8 also appears to be selected from the test-set sweep in Fig. 1.

(d) The "tail" metric hides a large head-class cost. Because "tail" covers 80% of classes, it is close to the overall average. From Table 1, AWP+AURORA on C10-LT (ResNet) implies head clean accuracy of 63.2% (vs. 96.6% for AWP base) and head PGD accuracy of 37.9% (vs. 85.3%), with tail ≥ head. TRADES+AURORA, under the same rule, keeps head PGD at 63.1%. The same allocation rule therefore produces qualitatively different head/tail behavior across base learners, which is unexplained. Many/medium/few splits and worst-class robust accuracy, as reported in the class-wise robustness literature (CFA; Li & Liu, 2023), are needed to judge the trade-off.

W2. The novelty is overstated relative to existing geometry-aware and class-wise reweighting

(a) FTR is a class-averaged GAIRAT geometry value. The paper's own Lemma 5 shows that the empirical path area is D^c=1Mc∑i(K−Ti+1)+/K\hat D_c = \frac{1}{M_c}\sum_i (K - T_i + 1)_+/K, i.e., the class mean of a normalized "least number of PGD steps needed to misclassify." This is exactly the geometry value κ\kappa used by GAIRAT (Zhang et al., ICLR 2021) to reweight adversarial losses, and the same step-count signal underlies FAT's early-stopped PGD (Zhang et al., ICML 2020). MAIL (Liu et al., NeurIPS 2021) later argued that step counts are a coarse, path-dependent proxy and replaced them with probabilistic margins. None of these works is cited or compared. Class-wise reweighting from robust error is also established in fair adversarial training (FRL, Xu et al., ICML 2021; WAT, Li & Liu, AAAI 2023), which is likewise missing. The novel element reduces to aggregating κ\kappa per class and wrapping it in a Wilson bound.

(b) URA reintroduces a frequency prior, and its statistical framing is mis-specified. NcN_c in Eq. (8) is the full class training count every epoch, so the Wilson premium (roughly zp^(1−p^)/Ncz\sqrt{\hat p(1-\hat p)/N_c}) is a deterministic, monotone function of class frequency. This contradicts the claim that URA avoids a "frequency-to-weight mapping." The effect is also small: on C10-LT (sizes 5000→100), equal raw failure rates map to Wilson weights between about 0.94 and 1.11, a change of under ±10% relative to raw-rate weighting. That is hard to reconcile with the +4.2 PGD(all) and +4.3 PGD(tail) improvement of URA over "Endpoint Failure" in Table 4 without seeds or variance. More fundamentally, the training-set failure count over the whole class is not a binomial sample of the quantity of interest: the relevant gap for tail classes is between training and test robust error (robust overfitting; Rice et al., ICML 2020), which tail classes suffer most and which a binomial interval on training failures does not address. Statistics are also accumulated under a changing θ\theta within the epoch.

(c) The theory does not support the method's claims. Propositions 1–5 are algebraic properties of exponentiation, clipping and mean normalization, and Lemma 1 is the textbook Wilson inversion. The appendix itself states they do not imply improved robustness. They should not be presented as a contribution.

Questions for authors

Clarifications (could change my assessment):

Q1. Please explain the AWP+RLT C10-LT entries (Tables 1 and 3) that imply head clean accuracy above 100%. Were baseline numbers re-run under your pipeline or taken from other papers? Which checkpoint (best-robust or last epoch) is reported for each method?

Q2. Why does AutoAttack give higher accuracy than PGD-20 for PGD-AT Base in all nine settings? Please confirm the evaluation code, ϵ\epsilon, and checkpoint are identical across columns.

Q3. Do the Base and plug-in columns (UDR, CFA, DAFA, RLT) use the same prior correction (τ=1\tau=1) and RLT perturbation schedule as AURORA? If not, please report either all baselines with the recipe or AURORA without it.

Q4. How was λ=0.8\lambda=0.8 selected? Fig. 1 appears to be on the test set; was a validation split used?

Q5. In Eq. (3), xi(0)=xix^{(0)}_i = x_i, but PGD-AT starts from a uniform random point and TRADES from small Gaussian noise. Is eligibility aia_i and first passage measured from the clean input or the random start? A flip at k=1k=1 may reflect the random initialization rather than the attack.

Q6. How does FTR differ, conceptually and empirically, from averaging GAIRAT's κ\kappa per class?

Experiments needed to support the claims:

Q7. Mean ± std over at least 3 seeds for Table 1 (C10-LT) and Table 4.

Q8. Gradient-masking checks: individual AutoAttack components (APGD-CE, APGD-T, FAB-T, Square), the logit-scaling PGD of Hitaj et al. (2021), and PGD with more steps and restarts, reporting the PGD-to-AA gap for AURORA and baselines.

Q9. Many/medium/few-shot and worst-class robust accuracy, plus per-class weight trajectories wc(t)w_c(t) and their correlation with log⁡nc\log n_c. Q10. Ablations under the identical recipe: (i) FTR on raw endpoint rates without URA; (ii) D^c\hat D_c without the Wilson operator; (iii) a GAIRAT-style class-mean κ\kappa; (iv) inverse-frequency or class-balanced weighting (Cui et al., CVPR 2019).

Q11. A comparison with Cho et al. (ICLR 2025), which is cited but not evaluated, and REAT/TAET on WRN. Results at the more common CIFAR-10-LT imbalance ratio of 100, and a justification for using ratio 10 on CIFAR-100-LT and Tiny-ImageNet-LT.

Suggestions for future work (not affecting my score): ImageNet-LT scale; analysis of how the allocation interacts with robust overfitting over training.

References -

Zhang, J., Zhu, J., Niu, G., Han, B., Sugiyama, M., Kankanhalli, M. Geometry-aware Instance-reweighted Adversarial Training (GAIRAT). ICLR 2021.

Zhang, J., Xu, X., Han, B., Niu, G., Cui, L., Sugiyama, M., Kankanhalli, M. Attacks Which Do Not Kill Training Make Adversarial Learning Stronger (FAT). ICML 2020.

Liu, F., Han, B., Liu, T., Gong, C., Niu, G., Zhou, M., Sugiyama, M. Probabilistic Margins for Instance Reweighting in Adversarial Training (MAIL). NeurIPS 2021.

Hitaj, D., Pagnotta, G., Masi, I., Mancini, L. V. Evaluating the Robustness of Geometry-Aware Instance-Reweighted Adversarial Training. arXiv:2103.01914, 2021.

Xu, H., Liu, X., Li, Y., Jain, A., Tang, J. To be Robust or to be Fair: Towards Fairness in Adversarial Training (FRL). ICML 2021.

Li, B., Liu, W. WAT: Improve the Worst-Class Robustness in Adversarial Training. AAAI 2023.

Rice, L., Wong, E., Kolter, J. Z. Overfitting in Adversarially Robust Deep Learning. ICML 2020.

Cui, Y., Jia, M., Lin, T.-Y., Song, Y., Belongie, S. Class-Balanced Loss Based on Effective Number of Samples. CVPR 2019.

Cho, S., Lee, H., Kim, C. Long-tailed Adversarial Training with Self-Distillation. ICLR 2025

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

This paper looks into the problem of unreasonable class-level robustness allocation in long-tail adversarial training and proposes the AURORA framework. The method uses URA to estimate class attack risk and its uncertainty via Wilson confidence bounds, and further adjusts the risk through FTR using the first-time failure in attack trajectories, dynamically tweaking the robustness loss weights for different classes. Tests on various CNN architectures across multiple datasets show that, compared to baselines, this approach significantly boosts adversarial robustness.

Critical strengths

This article points out that in long-tail adversarial training, class frequency and empirical attack failure rates are not enough to fully reflect the actual robustness needs, and it tries to combine uncertainty estimation from limited samples with attack trajectory information for dynamic risk allocation. Although similar techniques have been done before, this idea of class-level risk modeling still has some research value.

Critical weaknesses
  1. The paper reports significant performance improvements, but the supplementary code shows that AURORA and the AWP baseline differ in prior calibration, perturbation scheduling, and the adversarial sample generation process, making it hard for the current experiments to isolate the actual contributions of URA and FTR. Also, FTR’s trajectory prediction and endpoint failure detection use eval() and train() modes respectively, which might affect the consistency of the first-failure statistics, so the experiments aren’t fully reliable.

  2. It seems like some related work is missing, like GAIRAT [1], which uses the iteration number when misclassification first occurs during the PGD process as a proxy signal for adversarial vulnerability and guides training reweighting. In this paper, FTR seems to be similar, so the innovation of this paper seems a bit lacking.

  3. On the other hand, in terms of paper writing, I feel that the paper is overcomplicated and very difficult for me to read, and that the table design is also very hard to follow.

[1] Geometry-aware Instance-reweighted Adversarial Training.

Questions for authors
  1. Based on Weakness 1, provide a reasonable experiment and the final revised code.

  2. Explain in detail the differences between FTR and GAIRAT, and if possible, I hope you can do some experimental comparisons as well.

  3. On the other hand, in terms of paper writing, I feel that the paper is overcomplicated and very difficult for me to read, and that the table design is also very hard to follow.

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

Related papers

Discussion (0)

Sign in to comment.