acceptodds
Under review as a conference paper at ICLR 2027

Understanding Adversarial Robustness through Semantic-Region Membership

Abstract

Adversarial robustness remains an important research topic for improving the reliability and security of deep neural networks. Conventional research on adversarial attacks and defences mainly evaluates robustness through prediction accuracy on clean images and their corresponding adversarial examples. In this paper, we define class-specific semantic regions based on the geometric organisation of learned feature representations and conduct experiments across a range of adversarially robust models and a model trained only on clean images. Our results reveal that adversarial robustness is not associated with stronger semantic organisation of clean representations, but rather with the preservation of semantic-region membership under adversarial perturbations. we derive a sufficient condition to establish the relationship between semantic-region membership preservation and adversarial feature displacement. We employ a feature-space PGD attack to empirically estimate adversarial feature displacement and evaluate the derived sufficient condition for semantic-region membership preservation. The results show that all feature representations of adversarial examples lie within their corresponding semantic region if they satisfy that condition. Another interesting finding is that even an adversarial example has been misclassified, its feature representation may still lie within its corresponding semantic region. These findings extends the robustness studies from conventional prediction-level to semantic organisation in learn feature space. feature representation may still lie within its corresponding semantic region. These findings extend robustness studies from the prediction level to semantic organisation in the learned feature space.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Adversarial robustness, image classification, computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kALSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adversarial robustness from the perspective of feature-space semantic organization. It defines a class-specific semantic region according to the nearest class centroid and measures whether samples preserve their semantic-region membership after adversarial perturbation. The paper further derives a sufficient condition relating the semantic margin to adversarial feature displacement and approximates the displacement using feature-space PGD. Experiments on CIFAR-10 and CIFAR-100 with multiple robust models suggest that robust models preserve semantic-region membership substantially better than a standard model, and that some misclassified adversarial examples can still remain in their true-class semantic regions.

Critical strengths

The paper provides a coherent feature-space analysis of adversarial robustness and evaluates the proposed measurements across a relatively large collection of existing robust models. The observation that prediction failure does not always coincide with loss of true-class centroid membership is also potentially useful for understanding the distinction between representation robustness and final classifier decisions.

Critical weaknesses

The main conceptual contribution is limited. The proposed semantic region is essentially the Voronoi region induced by class centroids in the learned feature space, while centroid-based representations, nearest-centroid classification, and class geometry have been extensively studied. The paper mainly reinterprets this established geometric construction as a semantic region, and it is not sufficiently demonstrated that the resulting analysis provides fundamentally new understanding beyond existing feature-geometry and neural-collapse studies.

The theoretical and empirical evidence is also not strong enough to support the broader claims. The sufficient condition follows directly from the triangle inequality and is quite conservative; in many evaluated models only a small fraction of samples satisfy it despite much higher actual region-survival rates. Moreover, experiments are restricted to CIFAR-10/100 and norm-bounded image attacks, leaving unclear whether the observations generalize to larger datasets, architectures, or other adversarial settings.

Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Acknowledgment to All Reviewers

We thank all reviewers for your constructive feedback. We noticed that several of your concerns overlap; however, to make the rebuttal as convenient to read as possible for each of you, we have provided a clean and complete response for every reviewer individually.

We also sincerely apologise for the delayed submission. Our new adaptive attacks, along with the extensive ImageNet evaluation across multiple models and architectures, proved to be highly time-consuming. We appreciate your patience.

AuthorsOpenReview ↗

Response to Reviewer kALS (W1–W2)

W1. Conceptual novelty

As discussed in our response to Reviewer Fw25 (W2), we acknowledge that centroid-based representations and Voronoi regions are established geometric concepts. Our contribution is not the geometric construction itself, but the new understanding of adversarial robustness obtained through the proposed membership-based analysis.

Existing neural collapse studies (Papyan et al., 2020; Su et al., 2024) primarily characterise the global geometric organisation of features and class centroids. However, preserving global class geometry does not necessarily imply preserving individual semantic-region membership under perturbation. Our analysis reveals two important distinctions. First, stronger semantic organisation of clean representations does not necessarily mean stronger robustness. Instead, robustness is more closely associated with membership preservation under perturbation. Second, adversarial prediction failure does not necessarily mean the loss of true-class membership. These findings distinguish clean representation organisation, representation robustness, and prediction robustness as related but distinct properties.

Furthermore, our additional ImageNet experiments (Fw25 W3) reproduce these phenomena. We will update the manuscript to emphasize that our novelty lies in the membership-based characterisation of robustness and the resulting scientific insights, rather than the introduction of a new geometric construction.

W2. Theoretical and empirical evidence

We acknowledge that the sufficient condition follows from the triangle inequality. The purpose of deriving it is not to introduce a new mathematical inequality, but to explicitly establish the relationship between membership preservation, the clean semantic margin, and adversarial feature displacement.

Conservativeness. The condition is sufficient rather than necessary. Failure to satisfy it does not imply that the representation leaves its semantic region. Whether a representation crosses a region boundary depends on both the magnitude and direction of displacement, whereas our condition considers only magnitude. The discrepancy between condition satisfaction and actual survival is therefore expected. It highlights the conservativeness of a direction-independent sufficient condition.

Scope. As discussed in our response to Reviewer trmg (W1), the PGD-based estimate does not provide a certified upper bound on maximum displacement. Our analysis therefore characterises the relationship between semantic margins, observed displacements, and membership, rather than providing a robustness certificate.

Generalisation. To address the concern about experimental generality, we extended our evaluation to ImageNet using both CNN and Transformer architectures (Fw25 W3). The additional experiments reproduce the key phenomena observed on CIFAR-10 and CIFAR-100, including substantially higher survival rates for robust models, the discrepancy between condition satisfaction and observed preservation, and the retention of true-class membership despite misclassification. Although our empirical evaluation focuses on norm-bounded image perturbations, the geometric condition is not dependent on a specific attack algorithm. We will add a discussion of this distinction (the generality of the theoretical condition vs. the scope of our empirical findings) in the final version.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors’ efforts to clarify the conceptual contribution and the scope of the theoretical and empirical findings. While these clarifications are helpful, my main concerns regarding conceptual novelty and the strength of the theoretical contribution remain. Therefore, I will maintain my original score.

Reviewer_Fw25OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

Starting from the geometric organization of the learned feature space, this paper treats Voronoi-style nearest-centroid regions defined by class feature centroids as class-specific semantic regions, measures whether semantic membership is preserved under adversarial perturbations using the semantic margin and the conditional semantic-region survival rate, and compares multiple RobustBench robust models with a standard WRN on CIFAR-10/100. Combining PGD-20/AutoAttack, interpolation trajectories, a random-label baseline, centroid stability, feature-space PGD displacement estimation, and prediction–semantic region consistency analysis, it argues that adversarial robustness is more closely associated with the preservation of semantic-region membership under perturbation than with stronger semantic organization of clean representations, and that misclassification does not necessarily imply the loss of true-class semantic information; theoretically, it provides a sufficient condition under which the semantic region is preserved when the maximum feature displacement is less than half the semantic margin, and empirically validates this condition using feature-space PGD to estimate the displacement.

Critical strengths
  1. The paper extends adversarial robustness analysis from the perspective of semantic-region membership in feature space, proposing the semantic margin, the conditional semantic-region survival rate, and a checkable sufficient condition, which offers a certain degree of novelty.
  2. The paper finds that misclassification of adversarial examples does not necessarily imply the loss of true-class semantic information, and provides relatively systematic empirical validation through multiple models, interpolation analysis, and feature-space PGD.
Critical weaknesses
  1. Please discuss and justify why the penultimate-layer features are used to construct the class-specific semantic regions, rather than other layers or intermediate representations; the impact of this choice on the semantic-region definition, centroid stability, and reliability of the conclusions should be clarified.
  2. Optimization for class-space consistency and related concepts have been proposed previously. Please clarify the novelty of this work relative to existing work and provide a distinction and discussion with respect to related methods.
  3. The experiments are limited to CIFAR-10 and CIFAR-100, lacking validation on larger-scale datasets such as ImageNet; additional experiments are recommended to support the generality of the conclusions.
  4. The attack evaluation only covers PGD and AutoAttack. It is recommended to include more attack types and adaptive attacks for validation. In particular, since the proposed sufficient condition and empirical validation rely on PGD-based feature-displacement estimation, please clarify whether the conclusions hold only for PGD or have broader implications for general robustness evaluation.
  5. The references and compared methods mainly stop around 2024, lacking literature analysis and experimental comparison with works from the past two years (especially 2025). It is recommended to supplement the latest related work and baselines to enhance timeliness and comparative sufficiency.
Questions for authors
  1. Why were penultimate-layer features chosen to define class-specific semantic regions rather than other layers or intermediate representations? Would the reported semantic-region membership and robustness conclusions change under a different layer choice?
  2. How does this work differ technically from prior class-space/feature-space consistency optimization? Please specify the novel methodological contribution beyond the empirical observation that semantic-region preservation correlates with robustness.
  3. Do the authors claim that the conclusions generalize beyond CIFAR-10/100? If yes, what evidence supports this beyond the current experiments; if not, please state the scope explicitly.
  4. Since the sufficient condition and its empirical validation rely on PGD-based feature-displacement estimation, do the conclusions apply only to PGD-style attacks or to general adversarial perturbations? Please clarify whether the AutoAttack results provide independent support or share the same PGD-based estimation limitation.
8 comments
AuthorsOpenReview ↗

Acknowledgment to All Reviewers

We thank all reviewers for your constructive feedback. We noticed that several of your concerns overlap; however, to make the rebuttal as convenient to read as possible for each of you, we have provided a clean and complete response for every reviewer individually.

We also sincerely apologize for the delayed submission. Our new adaptive attacks, along with the extensive ImageNet evaluation across multiple models and architectures, proved to be highly time-consuming. We appreciate your patience.

AuthorsOpenReview ↗

Response to Reviewer Fw25 (Part 1: W1–W2)

W1. Choice of penultimate-layer features

We chose the penultimate layer because our study focuses on class-level semantic information in the learned feature space. Unlike earlier layers that may encode low-level visual information, the penultimate layer provides the features directly used by the final classification head. This choice is also consistent with neural collapse studies (Papyan et al., 2020), which investigate the geometric organisation of penultimate-layer features and their relationship with class centroids.

More importantly, this feature space lets us distinguish between the loss of true-class semantic-region membership and classification failure, which is a central objective of our study. While our framework can be applied to intermediate layers, different layers may exhibit different levels of class separation and centroid stability. Our conclusions therefore specifically concern the feature space directly used for the final decision, rather than assuming these phenomena occur throughout all layers. We will add a brief discussion clarifying this scope in the final version.

W2. Novelty and related work

We appreciate the chance to clarify the distinction between our work and existing studies on class-space consistency.

Previous consistency-based methods aim to improve adversarial robustness by optimising feature representations during training. In contrast, our objective is to understand adversarial robustness through the geometric organisation of features that have already been learned. Our framework does not introduce a new training objective or require retraining. Instead, it provides a unified way to characterise and compare models trained with different defence strategies.

As discussed in Section 2.3, existing studies on adversarial neural collapse (e.g., Su et al., 2024) investigate the global geometric organisation of class centroids and the alignment between clean and adversarial simplex ETF structures. However, preserving global class geometry does not necessarily mean every individual feature representation remains associated with its class under perturbation. Our work explicitly investigates this sample-level relationship by defining class-specific semantic regions and measuring individual membership.

Beyond the empirical observations, we introduce the semantic margin and the conditional semantic-region survival rate to characterise class-specific membership. We also derive a sufficient condition linking membership preservation to the clean semantic margin and adversarial feature displacement. Overall, these measurements provide a diagnostic framework for investigating robustness beyond prediction accuracy and aggregate class geometry.

Applying this framework reveals that stronger clean semantic organisation does not necessarily mean stronger adversarial robustness, and that misclassification does not necessarily imply the loss of true-class semantic membership. We observe these phenomena consistently on CIFAR-10, CIFAR-100, and ImageNet (W3). We do not claim that centroids or nearest-centroid geometry are new. Rather, our contribution is using them as a systematic framework to characterise robustness at the level of individual semantic-region membership. This distinction will be addressed in the revision.

AuthorsOpenReview ↗

Response to Reviewer Fw25 (Part 2: W3 — Generalisation to ImageNet)

W3. Generalisation to ImageNet

We extended our experiments to ImageNet to test the generality of our findings. We evaluate three pretrained models: MeanSparse (Swin-L) (Amini et al., 2024), an adversarially trained Wide ResNet-50-2 (Salman et al., 2020), and a standard Wide ResNet-50-2 (Zagoruyko and Komodakis, 2016). The robust models are from RobustBench (Croce et al., 2021), and the standard model uses TorchVision's IMAGENET1K_V2 weights (PyTorch Contributors). All models are evaluated without additional training.

For each random seed, we randomly select 100 samples per class from the ImageNet training set to construct the class centroids, resulting in 100,000 samples across 1,000 classes. To assess centroid stability, we construct reference centroids using a separate set of five training samples per class. For evaluation, we randomly select five samples per class from the validation set, resulting in 5,000 evaluation samples. All subsets are therefore class-balanced. We repeat the experiments using three random seeds (88, 89, and 90), with an ℓ∞\ell_\infty perturbation budget of 4/2554/255 for all adversarial evaluations.

The results are in Tables 2–8. We observe the same key phenomena as on CIFAR: (i) the standard model shows stronger clean semantic-region membership than robust models; (ii) robust models have substantially higher survival rates under perturbation; (iii) the condition satisfaction rate is much lower than the observed survival rate; and (iv) a proportion of misclassified adversarial examples remain in their true-class semantic regions, especially for robust models.

These consistent observations across three datasets of different scales, and across CNN and Transformer architectures, provide further empirical support for our central finding: adversarial robustness is more closely associated with the preservation of semantic-region membership under perturbation than with stronger semantic organisation on clean data alone.

Table 2. Conditional semantic-region survival rates on ImageNet (%)

ModelPGD-20AutoAttack
Amini et al. (2024)80.69 ± 1.2586.85 ± 1.20
Salman et al. (2020)62.43 ± 0.6874.92 ± 0.93
Standard WRN-50-23.03 ± 0.174.14 ± 0.06

Conditional semantic-region survival rates under PGD-20 and ℓ∞\ell_\infty AutoAttack on ImageNet (ϵ=4/255\epsilon=4/255). PsurviveP_{\mathrm{survive}} is the unweighted mean of class-wise survival rates, conditioned on m(x)>0m(x)>0. Classes with no eligible clean samples are excluded. Results are reported as mean ± standard deviation across three runs (seeds 88, 89, and 90).

Table 3. Empirical sufficient-condition analysis on ImageNet

ModelCondition Satisfied (%)Region Survival (%)Condition Violation (%)
Amini et al. (2024)30.53 ± 0.0692.69 ± 0.190.00 ± 0.00
Salman et al. (2020)1.56 ± 0.2184.31 ± 0.180.00 ± 0.00
Standard WRN-50-20.00 ± 0.003.20 ± 0.13N/A

Empirical sufficient-condition analysis on ImageNet under feature-space PGD (ϵ=4/255\epsilon=4/255, 20 steps, three restarts). Condition Satisfied denotes the proportion of samples with m(x)>0m(x)>0 satisfying D^(x,ϵ)<m(x)/2\hat{D}(x,\epsilon)<m(x)/2. Region Survival denotes the proportion of samples with m(x)>0m(x)>0 remaining in their true-class region. Condition Violation is the proportion of condition-satisfying samples leaving that region under the same perturbation. N/A indicates that no samples satisfy the condition. The displacement estimate is not a certified upper bound. Results are reported as mean ± standard deviation across three runs (seeds 88, 89, and 90).

Reviewer_trmgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes adversarial robustness through whether perturbed features remain closest to their true-class centroid. It introduces membership-preservation metrics and a sufficient geometric condition, finding that robust models better preserve membership—even when predictions are wrong. Validation uses CIFAR-10 and CIFAR-100, with 13 robust models and one standard model per dataset, repeated over three data-sampling seeds. Experiments include PGD-20 and AutoAttack, random-label controls, centroid stability checks, clean-to-adversarial interpolation, and feature-displacement attacks. Results support the reported membership patterns; however, the empirical displacement estimate does not certify worst-case robustness, and zero violations for the same perturbations satisfying the sufficient condition follow directly from the geometric inequality.

Critical strengths

1.Separating prediction failure from feature-level class information. The paper shows that some misclassified adversarial examples remain closest to their true-class centroid. This provides a useful diagnostic distinction between classification-head errors and changes in feature geometry, motivating analysis beyond final prediction accuracy.

2.Empirical coverage across diverse robust models. Evaluating 14 models per dataset on CIFAR-10 and CIFAR-100, alongside multiple attacks, centroid-stability checks, and random-label controls, strengthens the evidence that the observed membership-preservation patterns are not confined to a single model.

Critical weaknesses

1.The sufficient-condition experiment provides limited independent validation. The paper measures feature displacement for a perturbation and checks region preservation under that same perturbation. Once the measured displacement is below half the semantic margin, preservation follows directly from the triangle inequality, making the reported zero violations mathematically expected. Moreover, PGD supplies a lower bound on maximum displacement, not the upper bound needed for certification. This substantially limits the evidential value of a central validation experiment.

2.The attacks do not directly test semantic-region robustness. Classification attacks target the original prediction head, while the proposed regions correspond to a nearest-centroid classifier. Consequently, incorrect predictions with preserved region membership may reflect differences between decision boundaries rather than independently robust semantic information. Maximizing feature displacement does not resolve this mismatch, since large displacement need not cross a region boundary. Without attacks directly targeting region membership, the paper’s central robustness interpretation remains insufficiently supported.

Questions for authors

1.In the sufficient-condition experiment, is region membership evaluated using the same perturbation used to compute (\hat D)? If so, what evidence does the zero-violation result provide beyond the triangle inequality?

2.Have you evaluated attacks that directly minimize the semantic margin or target the nearest-centroid classifier? do they preserve the reported gap between robust and standard models?

2 comments
AuthorsOpenReview ↗

Acknowledgment to All Reviewers

We thank all reviewers for your constructive feedback. We noticed that several of your concerns overlap; however, to make the rebuttal as convenient to read as possible for each of you, we have provided a clean and complete response for every reviewer individually.

We also sincerely apologise for the delayed submission. Our new adaptive attacks, along with the extensive ImageNet evaluation across multiple models and architectures, proved to be highly time-consuming. We appreciate your patience.

AuthorsOpenReview ↗

Response to Reviewer trmg

W1. Empirical validation of the sufficient condition

We thank the reviewer for distinguishing between empirical analysis and independent validation. We confirm that membership is evaluated using the same perturbation used to compute D^(x,ϵ)\hat{D}(x,\epsilon).

We completely agree that zero violations are mathematically expected when the measured displacement satisfies the condition. As the reviewer correctly points out, this is a sanity check of our implementation, not independent validation of the proposition. We also agree that the PGD-based estimate satisfies D^(x,ϵ)≤D(x,ϵ)\hat{D}(x,\epsilon)\leq D(x,\epsilon), giving a lower bound rather than a certified upper bound. Satisfying D^(x,ϵ)<m(x)/2\hat{D}(x,\epsilon)<m(x)/2 does not guarantee preservation for all perturbations within the budget. We acknowledge this limitation in Section 3.5.

The key finding from this experiment is not the absence of violations, but the relationship between the clean semantic margin, observed displacement, and membership preservation across models. Robust models show higher condition satisfaction rates than the standard model under feature-displacement attacks. Furthermore, their observed survival rates are much higher than their condition satisfaction rates, which illustrates the conservativeness of the sufficient condition. We reproduce these observations on ImageNet in our response to Reviewer Fw25 (W3).

To avoid misunderstanding, we will rename Section 3.5 from "Empirical Validation of the Sufficient Condition" to "Empirical Analysis of Semantic Margin and Adversarial Feature Displacement." We will also clarify that the zero-violation result is a consistency check, not a theoretical validation or robustness certificate.

W2. Direct semantic-region attacks

To directly address this concern, we additionally evaluate a semantic-region-targeted PGD (Madry et al., 2018) attack that explicitly minimises the semantic margin m(x+δ)m(x+\delta) instead of attacking the original classification head. We use 20 optimisation steps and three random restarts, with the same ℓ∞\ell_\infty perturbation budgets as in the classification-based evaluations (8/2558/255 for CIFAR-10 and CIFAR-100 and 4/2554/255 for ImageNet).

As shown in Table 1, directly targeting semantic-region membership generally reduces the survival rates compared with classification-based PGD and AutoAttack. However, the substantial gap between robust and standard models remains. On CIFAR-10, Mean Sparse and Hendrycks achieve survival rates of 80.80% and 61.43%, respectively, compared with 0.00% for the standard model. On CIFAR-100, the corresponding rates are 59.66%, 42.26% and 0.00%. We observe the same phenomenon on ImageNet, where Mean Sparse and Salman achieve 76.31% and 54.27%, respectively, compared with 0.02% for the standard model.

These results suggest that the observed difference in semantic-region preservation between robust and standard models is not solely attributable to the mismatch between the original classification boundaries and the nearest-centroid boundaries. The relative ranking of the evaluated models also remains unchanged under this direct attack across all three datasets. Here, semantic-region membership is an operational measure of class association in the learned feature space, rather than a claim that nearest-centroid membership captures all human-perceived semantic information.

Table 1. Conditional semantic-region survival rates (%)

DatasetModelPGD-20AutoAttackSemantic PGD
CIFAR-10Amini et al. (2024)85.52 ± 1.2089.24 ± 0.4980.80 ± 1.72
CIFAR-10Hendrycks et al. (2019)66.93 ± 0.5272.72 ± 0.2061.43 ± 0.30
CIFAR-10Standard0.00 ± 0.000.07 ± 0.050.00 ± 0.00
CIFAR-100Amini et al. (2024)70.16 ± 1.1677.34 ± 1.0059.66 ± 1.66
CIFAR-100Hendrycks et al. (2019)54.68 ± 1.4866.20 ± 1.5742.26 ± 0.53
CIFAR-100Standard0.38 ± 0.380.50 ± 0.270.00 ± 0.00
ImageNetAmini et al. (2024)80.69 ± 1.2586.85 ± 1.2076.31 ± 1.26
ImageNetSalman et al. (2020)62.43 ± 0.6874.92 ± 0.9354.27 ± 0.47
ImageNetStandard3.03 ± 0.174.14 ± 0.060.02 ± 0.02

Survival rates are conditioned on m(x)>0m(x)>0 and averaged over classes with eligible clean samples. The perturbation budget is 8/2558/255 for CIFAR-10 and CIFAR-100 and 4/2554/255 for ImageNet. Semantic PGD uses 20 steps and three random restarts. Results are mean ± standard deviation across seeds 88, 89 and 90. Bold indicates the highest survival rate within each dataset and attack.

Comments on the paper

1 comment
AuthorsOpenReview ↗

References for Author Rebuttal

The following references are shared across our responses to all reviewers (trmg, Fw25, and kALS).

References

  • Sajjad Amini, Mohammadreza Teymoorianfard, Shiqing Ma, and Amir Houmansadr. MeanSparse: Post-training robustness enhancement through mean-centered feature sparsification. arXiv preprint arXiv:2406.05927, 2024.

  • Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, 2020.

  • Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. RobustBench: A standardized adversarial robustness benchmark. In NeurIPS Datasets and Benchmarks Track, 2021.

  • Dan Hendrycks, Kimin Lee, and Mantas Mazeika. Using pre-training can improve model robustness and uncertainty. In International Conference on Machine Learning, 2019.

  • Binghui Li and Yuanzhi Li. Adversarial training can provably improve robustness: Theoretical analysis of feature learning process under structured data. In International Conference on Learning Representations, 2025.

  • Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018.

  • Vardan Papyan, X. Y. Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020. doi: 10.1073/pnas.2015509117.

  • PyTorch Contributors. TorchVision: Wide resnet-50-2 pretrained weights. https://docs.pytorch.org/vision/main/models/generated/torchvision.models.wide_resnet50_2.html. IMAGENET1K_V2 weights.

  • Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust ImageNet models transfer better? In Advances in Neural Information Processing Systems, volume 33, pages 3533–3545, 2020.

  • Jingtong Su, Ya Shi Zhang, Nikolaos Tsilivis, and Julia Kempe. On the robustness of neural collapse and the neural collapse of robustness. Transactions on Machine Learning Research, 2024.

  • Zeming Wei, Steven Y. Guo, and Yisen Wang. Identifying and understanding cross-class features in adversarial training. In International Conference on Machine Learning, volume 267, pages 65997–66014, 2025.

  • Yibo Xu, Dawei Zhou, Decheng Liu, Nannan Wang, and Xinbo Gao. Fst: Improving adversarial robustness via feature similarity-based targeted adversarial training. Pattern Recognition, 175:113010, 2026. doi: 10.1016/j.patcog.2025.113010.

  • Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, 2016.

  • Kejia Zhang, Juanjuan Weng, Shaozi Li, and Zhiming Luo. Towards adversarial robustness via debiased high-confidence logit alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783–2792, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.