acceptodds
Under review as a conference paper at ICLR 2027

Attack-Free Adversarial Vulnerability Assessment via Second-Order Statistics of Predictive Probabilities

Abstract

Deep neural networks are widely deployed in safety-critical applications such as medical image analysis and autonomous driving, where adversarial robustness is essential. Existing robustness evaluations primarily measure overall performance degradation under attack, offering limited insight into class-wise or directional vulnerability patterns. We propose a second-order statistical framework that treats class-wise softmax outputs as random variables and analyzes their conditional variance–covariance structure to characterize structural vulnerability. Using only clean validation data-without generating adversarial examples or relying on specific attack algorithms-the framework enables attack-agnostic evaluation while significantly reducing computational cost, serving as a practical pre-screening tool for rapid model assessment. The conditional variance of the correct-class probability quantifies relative class-wise adversarial vulnerability, while conditional covariance reveals directional misclassification tendencies. Experiments on standard benchmarks under diverse untargeted and targeted attack settings demonstrate strong alignment between the proposed metrics and class-wise attack success rates, suggesting that adversarial vulnerability can be effectively characterized without explicit attack generation.

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

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_Vo1KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a class-wise adversarial vulnerability assessment framework based solely on clean data. By analyzing the softmax prediction probabilities of a model on a clean dataset, the framework aims to estimate the adversarial vulnerability of different classes. Specifically, the authors introduce two metrics, SOIV-U and SOIV-T. SOIV-U uses the ideal-centered second-order moment of the true-class prediction probability to assess class-wise vulnerability under untargeted attacks, while SOIV-T employs the cross second order moment of source class and target class prediction probabilities to characterize directional vulnerability under targeted attacks. Experimentally, the authors evaluate the proposed framework across multiple datasets using various adversarial attack methods and report correlations between the proposed metrics and the observed class-wise ASRs.

Critical strengths

The SOIV framework proposed in this paper only relies on the Softmax prediction probabilities on a clean validation set to obtain class-level and source-to-target directional vulnerability statistics. The authors report about a 162x speed-up in computation under specific experimental settings. This low-cost evaluation method is suitable for model pre-screening scenarios where you need to quickly identify potentially vulnerable classes.

Critical weaknesses
  1. Some relevant literature is missing. Existing studies have separately explored robustness evaluation without attacks [1], class level differences in adversarial robustness and adaptive defenses [2], as well as source-to-target class attack tendencies [3]. Although this paper proposes specific second order probabilistic statistical metrics, their incremental contribution compared to these existing studies has not been fully clarified. In addition, SOIV-T lacks a direct comparison with simple source-to-target conditional probability statistics, so the independent value of its directional vulnerability estimation is still unclear.

  2. This paper mainly verifies the effectiveness of the SOIV and ASR correlation methods, but in the targeted attack experiments, only one source class is chosen for each dataset and model, making it hard to prove that the proposed metrics can accurately capture the full vulnerability relationships between classes. Also, the authors choose the attack intensity based on the class-level ASR distribution, and Appendix D.7 shows that the correlation changes noticeably with attack intensity. So, the current experiments aren’t enough to show that SOIV can provide a stable and reliable vulnerability assessment under different attack conditions.

[1] Attack-free evaluating and enhancing adversarial robustness on categorical data. [2] To be Robust or to be Fair: Towards Fairness in Adversarial Training. [3] Improving robust fairness via balance adversarial training.

Questions for authors
  1. In Section 3, the author estimates adversarial vulnerability through statistics of predicted probabilities on clean samples. However, SOIV-U includes both prediction variance and confidence bias, while SOIV-T is closely related to category confusion of normal samples. Could the author further explain why these metrics can reflect adversarial vulnerability, rather than just normal classification difficulty, and demonstrate their added value compared to simple confidence statistics?

  2. In Section 4.2, the author only validates SOIV-T using Automobile as the source category, yet in Section 4.3, they present a complete source-target vulnerability matrix. Could the author explain why this experiment is enough to support conclusions about the overall directional vulnerability structure, and provide validation results for more source categories?

  3. In Section 4.1, the author verifies the effectiveness of SOIV through its correlation with various attack ASRs. However, different attacks’ optimization strengths and perturbation constraints might affect ASR and class ranking. Could the author clarify whether the attack settings used were fully optimized, and verify if SOIV can still maintain stable correlation under a unified threat model and more fully optimized attack conditions?

5 comments
AuthorsOpenReview ↗

Response to Weakness 1 & Question 1 (Novelty)

We thank the reviewer for highlighting [1–3], which study related but distinct aspects of adversarial robustness. Specifically, [1] introduces gradient-based IGSG to evaluate feature sensitivity on categorical data. [2] examines class-wise robustness disparities arising from adversarial training and proposes a fairness-oriented method. [3] analyzes source-class attack difficulty and target-class bias to improve adversarial training. In contrast, our work estimates class-wise and source-to-target vulnerability in image classifiers using only clean softmax outputs, without generating adversarial examples or computing input gradients. Our contribution is not the discovery of class imbalance or attack directionality, but an attack-free approach to characterizing their relative vulnerability structure before performing attacks.

The adversarial relevance of SOIV-U follows from its ideal-centered formulation

E[(1−ps)2∣Y=s]=Var(ps∣Y=s)+(1−E[ps∣Y=s])2.\mathbb{E}[(1-p_s)^2\mid Y=s]=\mathrm{Var}(p_s\mid Y=s)+(1-\mathbb{E}[p_s\mid Y=s])^2.

Thus, SOIV-U captures both within-class confidence variability and the squared mean confidence deficit. Section 3.4 and Appendix A.1 provide a conditional theoretical motivation for its relationship with adversarial vulnerability. Under local smoothness and sufficient coverage of probability-sensitive directions by natural within-class variations, clean probability variability and small-perturbation sensitivity depend on the same local probability gradient. This supports SOIV-U as a relative vulnerability indicator without implying equivalence to ASR or a robustness guarantee.

Ideal centering also addresses a limitation of conventional sample variance. Since variance is centered at each class's empirical mean, it can become small even when true-class probabilities are consistently close to zero. In contrast, SOIV-U measures deviations from the ideal confidence of one. Appendix D.4 demonstrates this distinction experimentally on adversarially trained CIFAR-100 models. Sample variance exhibits negative correlations with ASR across untargeted attacks (Tables 32 and 34), whereas SOIV-U maintains positive correlations (Tables 19 and 21). This provides a concrete advantage in assessing adversarial vulnerability beyond simply using a different statistical formulation.

We further compare SOIV-U with simpler statistics on CIFAR-10 using ResNet-18. All statistics are computed from 5,000 clean validation images, while ASR is evaluated on a separate test set of 10,000 images. The MIXED untargeted setting aggregates FGSM, PGD, C&W, Square, and AutoAttack (Section 4.1, Table 1). Here, Variance denotes the unbiased within-class variance of true-class probabilities, while Confidence is defined as 1−E[ps∣Y=s]1-\mathbb{E}[p_s\mid Y=s] so that larger values indicate greater vulnerability.

CorrelationSOIV-UVarianceConfidence
Pearson0.927 (1.1e-04)0.899 (4.0e-04)0.843 (2.2e-03)
Spearman0.915 (2.1e-04)0.915 (2.1e-04)0.879 (8.1e-04)
Kendall0.778 (9.5e-04)0.733 (2.2e-03)0.689 (4.7e-03)

Parentheses indicate two-sided p-values. SOIV-U matches or exceeds both baselines here, without implying universal predictive superiority.

For directional vulnerability, the magnitude of SOIV-T is

E[(1−ps)pt∣Y=s].\mathbb{E}[(1-p_s)p_t\mid Y=s].

SOIV-T measures directional vulnerability by jointly considering the source-confidence deficit and target probability, ∣SOIV-T(s,t)∣=E[(1−ps)pt∣Y=s]|\mathrm{SOIV\text{-}T}(s,t)|=\mathbb{E}[(1-p_s)p_t\mid Y=s], whereas conditional target probability, E[pt∣Y=s]\mathbb{E}[p_t\mid Y=s], uses only the latter. We compare both metrics on CIFAR-10/ResNet-18 using airplane as the source class and MIXED targeted ASR from FGSM, PGD, and C&W.

MetricPearsonSpearmanKendall
Conditional Target Probability0.64200.70000.5000
Absolute SOIV-T0.72640.85000.6667

SOIV-T achieves higher correlations across all three measures, particularly for Spearman and Kendall. These results suggest that incorporating the source-confidence deficit can improve alignment with targeted ASR in this setting. The second-order formulation also supports the unified variance–covariance analysis in Section 4.3.

Finally, Section 5.2 (Table 3) demonstrates practical relevance to defense. SOIV-U-guided adversarial training allocates adversarial examples by estimated class-wise vulnerability rather than uniformly. Under the same budget, overall AutoAttack ASR falls from 61.43% to 58.17% (L∞L_\infty, ϵ=8/255\epsilon=8/255), improving 7 of 10 classes. This single setting does not establish superiority over [2,3], but shows how SOIV estimates can guide defense resource allocation. Overall, SOIV offers an efficient, attack-free structural diagnostic rather than universally superior ASR prediction.

AuthorsOpenReview ↗

Response to Question 2 (Source-Class Validation)

We agree that evaluating only Automobile was insufficient to establish generality. We have therefore extended the targeted evaluation to the remaining nine CIFAR-10 source classes. The table reports the correlation between ∣SOIV-T∣|\mathrm{SOIV\text{-}T}| and targeted ASR across the nine possible target classes for each source. Each entry represents Pearson/Spearman/Kendall.

SourceFGSMPGDCWMIXED
Airplane0.701/0.783/0.5560.675/0.817/0.6110.719/0.817/0.6110.726/0.850/0.667
Bird0.647/0.661/0.5350.757/0.667/0.5560.762/0.633/0.5000.743/0.650/0.556
Cat0.575/0.767/0.6670.650/0.800/0.7220.771/0.733/0.6670.675/0.800/0.722
Deer0.549/0.717/0.5560.707/0.750/0.5560.679/0.717/0.5560.654/0.700/0.500
Dog0.801/0.895/0.7610.697/0.900/0.7780.832/0.850/0.7220.788/0.900/0.778
Frog0.804/0.733/0.5560.739/0.800/0.6670.796/0.700/0.5000.803/0.733/0.556
Horse0.823/0.900/0.7780.829/0.850/0.7220.821/0.900/0.7780.826/0.900/0.778
Ship0.935/0.613/0.5140.779/0.733/0.6110.889/0.617/0.5560.881/0.733/0.667
Truck0.902/0.750/0.5560.918/0.800/0.6670.958/0.833/0.6670.937/0.817/0.722

All correlations are positive across the additional source classes. Under MIXED attacks, Pearson ranges from 0.654 to 0.937 and Spearman from 0.650 to 0.900. These results support directional alignment beyond Automobile, although its strength varies by source class.

AuthorsOpenReview ↗

Response to Question 3 (Attack Settings)

We thank the reviewer for raising concerns about attack strength and optimization. The manuscript already provides evidence across attacks in Section 4.1 (Table 1), where SOIV-U is compared with FGSM, PGD, C&W, Square, AutoAttack, and MIXED. In particular, standard AutoAttack (L∞L_\infty, ϵ=1.5/255\epsilon=1.5/255) yields Pearson, Spearman, and Kendall correlations of 0.933, 0.915, and 0.778, respectively. Also, MIXED yields 0.927, 0.915, and 0.778. Appendix B.5 (Table 6) specifies the attack configurations, and all ASRs are measured on clean-correct test samples. Thus, the original evaluation already includes a standardized AutoAttack assessment rather than relying only on individually configured attacks.

The manuscript also examines why attack strength affects correlation. Section 3.4 and Appendix A.1 motivate SOIV through a conditional connection between clean probability variability and local adversarial sensitivity, based on a first-order approximation that may become less accurate for larger perturbations. Appendix D.7 (Table 39) explicitly examines stronger attacks and the spread of class-wise ASRs. As the PGD perturbation budget increases from ϵ=2/255\epsilon=2/255 to 4/2554/255, 8/2558/255, and 16/25516/255, the standard deviation of class-wise ASR decreases from 14.46 to 5.95, 1.01, and 0 percentage points, respectively. At the strongest PGD and C&W settings, all classes reach 100% ASR, making the correlation undefined because no meaningful class-wise ranking remains. This does not necessarily indicate instability in SOIV itself. For a fixed model and clean validation set, SOIV values remain unchanged, whereas the attack-based reference becomes saturated and loses its ability to distinguish vulnerable classes. However, we acknowledge that saturation does not explain every decline in correlation, since stronger attacks may also change relative class rankings. As explained in Appendix B.5, attack budgets were therefore selected to preserve measurable class-wise ASR differences, allowing meaningful evaluation of relative vulnerability.

To test whether the observed alignment depends on the original hyperparameters, we additionally evaluated CIFAR-10/ResNet-18 under a unified L∞L_\infty budget of ϵ=1.5/255\epsilon=1.5/255 and varied PGD optimization steps and perturbation budgets

Attack (ϵ\epsilon, steps)PearsonSpearmanKendall
FGSM (1.5/255)0.9590.9390.822
Square (1.5/255)0.9630.9520.867
AutoAttack (1.5/255)0.9330.9150.778
PGD (1.5/255, 20)0.9350.9150.778
PGD (1.5/255, 50)0.9370.9150.778
PGD (1.5/255, 100)0.9370.9150.778
PGD (2.5/255, 10)0.8240.8550.733
PGD (4/255, 10)0.7280.8060.689

The original six-step PGD setting (ϵ=1.5/255\epsilon=1.5/255, Table 1) achieved Pearson, Spearman, and Kendall correlations of 0.936, 0.939, and 0.822, respectively. Increasing PGD to 20–100 steps under the same perturbation budget produced broadly similar correlations, while larger budgets reduced the correlations without reversing their signs in the additional experiments. Nevertheless, increasing the number of PGD steps does not guarantee global attack optimality. Furthermore, Appendix D.8 (Table 40) evaluates nine combinations of data split and training seeds, reporting mean SOIV-U correlations of 0.870, 0.888, and 0.738 for Pearson, Spearman, and Kendall against MIXED ASR, respectively. These results provide additional evidence of reproducibility across random seeds, although they do not establish invariance across threat models. Overall, SOIV-U remains useful for relative class-wise vulnerability assessment when meaningful ASR differences exist, rather than guaranteeing consistent rankings under arbitrary attack strengths or saturated attack conditions.

Reviewer_8DH1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes two attack-free adversarial vulnerability metrics: SOIV-U for class-wise untargeted vulnerability and SOIV-T for source-to-target directional vulnerability, they are computed solely from clean softmax outputs. The paper empirically show that these metrics correlate with class-wise and targeted attack success rates across multiple attacks, datasets, and architectures, including ResNet, VGG, Swin, and ViT models. They further demonstrate that SOIV can provide substantial computational savings over attack-based evaluation and can be used to guide class-aware adversarial training.

Critical strengths
  1. SOIV in this paper is lightweight and practical: requires only clean softmax outputs, avoids adversarial-example generation, and provides both class-wise and directional vulnerability estimates at very low computational cost.
  2. The paper evaluates the proposed metrics across multiple attacks, datasets, architectures, calibration settings, and random seeds, and consistently finds meaningful correlation with adversarial vulnerability.
Critical weaknesses
  1. The paper does not adequately discuss or experimentally compare with highly relevant prior methods such as ROBY in [1] and RDI in [2], which also estimate adversarial vulnerability from clean/model statistics without generating adversarial examples. As a result, it is difficult to determine whether SOIV provides a genuinely distinct contribution or mainly an output-space variant of an existing line of work.

[1] Haibo Jin, Jinyin Chen, Haibin Zheng, Zhen Wang, Jun Xiao, Shanqing Yu, and Zhaoyan Ming. ROBY: Evaluating the adversarial robustness of a deep model by its decision boundaries. Information Sciences, 587:97–122, 2022. DOI:10.1016/j.ins.2021.12.021

[2] Jialei Song, Xingquan Zuo, Feiyang Wang, Hai Huang, and Tianle Zhang. RDI: An adversarial robustness evaluation metric for deep neural networks based on model statistical features. In Proceedings of the Forty-first Conference on Uncertainty in Artificial Intelligence (UAI 2025), PMLR 286:3999–4012, 2025.

  1. The proposed second-order formulation appears largely heuristic rather than theoretically necessary. In particular, SOIV-U is essentially an ideal-centered squared confidence error, and its effectiveness may come largely from incorporating the mean confidence deficit rather than from genuinely new second-order robustness information. The paper provides empirical motivation for this design, but does not establish why this particular quadratic formulation should be preferred over simpler confidence-based, margin-based, or proper-scoring-rule-based alternatives.
Questions for authors
  1. For SOIV-U, which component is primarily responsible for its correlation with adversarial vulnerability? the variance term or the mean-confidence deficit term? Since SOIV-U combines both, it would be helpful to clarify whether the second-order variability itself is essential to the observed effect.

  2. How much additional information does SOIV-T provide over simpler directional quantities, such as E[pt∣Y=s]\mathbb{E}[p_t\mid Y=s], mean source-target margins, or clean confusion tendencies? In particular, what aspect of directional vulnerability is captured specifically by the proposed cross-moment?

  3. Given its dependence on softmax confidence and calibration, should it primarily be interpreted as an intra-model class-wise pre-screening tool, or do the authors also intend it to support robustness comparison across different models?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 (Comparison with ROBY and RDI)

We thank the reviewer for suggesting ROBY and RDI. We will cite both and add them to the comparison tables in the revision. Both characterize the geometry of class distributions in feature space, whereas SOIV is computed from the predictive probability distribution. Since both aggregate into a single model-level score, we use the per-class quantities computed before their final aggregation.

The following table compares SOIV-U with ROBY and RDI under MIXED untargeted ASR for CIFAR-10 + ResNet-18, under the same setting as Table 26 in the appendix. All metrics are computed on the same 500-image subset of 50 images per class, and we report the mean and standard deviation over three subsampling seeds.

CorrelationSOIV-UROBYRDI
Pearson0.883 ± 0.0350.846 ± 0.0220.835 ± 0.018
Spearman0.899 ± 0.0190.842 ± 0.0420.859 ± 0.019
Kendall0.778 ± 0.0440.704 ± 0.0680.719 ± 0.026

Averaged over the three seeds, SOIV-U attains higher correlations than ROBY and RDI on all coefficients.

The following tables compare SOIV-U with ROBY and RDI for the adversarially trained PreActResNet-18 and ResNet-18 respectively, under the same setting as Tables 28 and 29 in the appendix. Here the validation set contains 50 images per class, so all metrics are computed on the full validation set without subsampling. The values in parentheses indicate p-values.

CorrelationSOIV-UROBYRDI
Pearson0.8552 (1.0e-29)0.4751 (5.9e-07)0.4311 (7.6e-06)
Spearman0.8810 (1.3e-33)0.4936 (1.8e-07)0.4390 (4.9e-06)
Kendall0.6971 (9.2e-25)0.3354 (7.7e-07)0.2914 (1.8e-05)
CorrelationSOIV-UROBYRDI
Pearson0.8608 (1.7e-30)0.5400 (6.7e-09)0.6036 (3.0e-11)
Spearman0.8867 (1.3e-34)0.5549 (2.1e-09)0.6153 (9.7e-12)
Kendall0.6981 (7.9e-25)0.3889 (9.9e-09)0.4277 (2.9e-10)

Unlike in the CIFAR-10 + ResNet-18 setting, ROBY and RDI yield notably lower correlations here, while SOIV-U and the uncertainty-related metrics retain correlations comparable to the CIFAR-10 setting, as reported in Tables 28 and 29. ROBY and RDI are computed from distances in feature space, while SOIV and the other attack-free metrics are computed from the predictive distribution, so the drop is confined to the distance-based metrics. This may be due to the larger number of classes or the adversarially trained model.

The main difference lies in what the metrics provide. RDI computes one value per class, and ROBY's pairwise term is symmetric, so neither can distinguish a source-to-target direction. In SOIV, the class-wise SOIV-U and the directional SOIV-T come from the same second-order object, and the unified SOIV matrix presents both together.

AuthorsOpenReview ↗

Response to Weakness 2 & Question 1 (Decomposition of SOIV-U)

We thank the reviewer for raising this important question regarding the contribution of second-order variability relative to the mean-confidence deficit.

The adversarial relevance of SOIV-U follows from its ideal-centered formulation:

E[(1−ps)2∣Y=s]=Var(ps∣Y=s)+(1−E[ps∣Y=s])2.\mathbb{E}[(1-p_s)^2\mid Y=s]=\mathrm{Var}(p_s\mid Y=s)+(1-\mathbb{E}[p_s\mid Y=s])^2.

Thus, SOIV-U captures both within-class confidence variability and the squared mean-confidence deficit. Section 3.4 and Appendix A.1 provide a conditional theoretical motivation for its relationship with adversarial vulnerability. Under local smoothness and sufficient coverage of probability-sensitive directions by natural within-class variations, clean probability variability and small-perturbation sensitivity depend on the same local probability gradient. This supports SOIV-U as a relative vulnerability indicator without implying equivalence to ASR or a robustness guarantee.

Ideal centering also addresses a limitation of conventional variance. Since variance is centered at each class's empirical mean, it can become small even when true-class probabilities are consistently close to zero. In contrast, SOIV-U measures deviations from the ideal confidence of one. Appendix D.4 demonstrates this distinction experimentally on adversarially trained CIFAR-100 models. Sample variance exhibits negative correlations with ASR across untargeted attacks (Tables 32 and 34), whereas SOIV-U maintains positive correlations (Tables 19 and 21). This provides a concrete advantage in assessing adversarial vulnerability beyond simply using a different statistical formulation.

We further compare SOIV-U with simpler statistics on CIFAR-10 using ResNet-18. All statistics are computed from 5,000 clean validation images, while ASR is evaluated on a separate test set of 10,000 images. The MIXED untargeted setting aggregates FGSM, PGD, C&W, Square, and AutoAttack (Section 4.1, Table 1). Here, Variance denotes the unbiased within-class sample variance of true-class probabilities, Squared Confidence Deficit is defined as (1−E[ps∣Y=s])2(1-\mathbb{E}[p_s\mid Y=s])^2, and Confidence Deficit is defined as 1−E[ps∣Y=s]1-\mathbb{E}[p_s\mid Y=s], so that larger values indicate greater vulnerability.

CorrelationSOIV-U (A+B)Variance (A)Squared Confidence Deficit (B)Confidence Deficit
Pearson0.927 (1.1e-04)0.899 (4.0e-04)0.830 (2.9e-03)0.843 (2.2e-03)
Spearman0.915 (2.1e-04)0.915 (2.1e-04)0.879 (8.1e-04)0.879 (8.1e-04)
Kendall0.778 (9.5e-04)0.733 (2.2e-03)0.689 (4.7e-03)0.689 (4.7e-03)

The variance component (A) exhibits stronger correlations with ASR than the squared mean-confidence deficit (B) across all three measures, indicating that within-class variability contributes meaningful information beyond the mean-confidence deficit alone. Moreover, SOIV-U (A+B) achieves higher Pearson and Kendall correlations than either component individually, while matching the variance component in Spearman correlation. SOIV-U also consistently outperforms the first-order confidence deficit in this setting.

Overall, these results provide empirical evidence that second-order variability contributes meaningful information beyond first-order confidence, while ideal centering integrates both variability and confidence deficit into a unified vulnerability indicator. We do not claim that this quadratic formulation is theoretically necessary or universally superior to simpler alternatives.

AuthorsOpenReview ↗

Response to Question 2 (Comparison with Simpler Directional Quantities)

We thank the reviewer for this question. SOIV-T measures directional vulnerability through the cross-moment ∣SOIV-T(s,t)∣=E[(1−ps)pt∣Y=s]|\mathrm{SOIV\text{-}T}(s,t)|=\mathbb{E}[(1-p_s)p_t\mid Y=s]. Unlike conditional target probability, E[pt∣Y=s]\mathbb{E}[p_t\mid Y=s], SOIV-T jointly considers the source-confidence deficit and target probability at the sample level. It therefore captures whether high target probability coincides with reduced source confidence, providing directional information beyond average target probabilities (Section 3.3 and Appendix A.2).

We further compare SOIV-T with mean source–target probability margin and clean confusion. For a fixed source class, the mean probability margin provides the same target-ranking information as conditional target probability after aligning its sign with vulnerability. Clean confusion captures observed misclassifications but does not distinguish targets with no clean misclassifications. In contrast, SOIV-T uses continuous probability information to distinguish source–target relationships even without observed clean misclassifications.

We evaluate these metrics on CIFAR-10/ResNet-18 using airplane as the source class and MIXED targeted ASR aggregated from FGSM, PGD, and C&W. Absolute correlation values are used for SOIV-T and mean source–target probability margin, since smaller values indicate greater vulnerability for these metrics. The values in parentheses indicate p-values.

MetricPearsonSpearmanKendall
SOIV-T0.7264 (2.7e-02)0.8500 (3.7e-03)0.6667 (1.3e-02)
Conditional Target Probability0.6420 (6.2e-02)0.7000 (3.6e-02)0.5000 (7.5e-02)
Mean Probability Margin0.6420 (6.2e-02)0.7000 (3.6e-02)0.5000 (7.5e-02)
Clean Confusion0.5586 (1.2e-01)0.6667 (5.0e-02)0.5512 (4.4e-02)

SOIV-T achieves the highest correlations across all three measures, particularly improving Spearman and Kendall correlations. These results suggest that jointly considering source-confidence deficits and target probabilities provides useful additional information for identifying relative directional vulnerability in this setting. Furthermore, SOIV-T forms the off-diagonal entries of the unified variance–covariance matrix in Section 4.3, enabling integrated analysis of class-wise and directional vulnerabilities using only clean validation outputs.

Reviewer_jfH3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a metric that computes the variance of a r.v., which is defined as the probability that DNN classify input to the correct class minus one. This quantity measures how far the true-class probability deviates from perfect confidence under clean inputs, denotes as the ideal-centered conditional variance, which is used to characterize probability of model is vulnerable to untargeted attacks. The paper also introduces the ideal-centered conditional covariance, which is intended to analyze whether a decrease in the probability for the correct class is associated with an increase in the probability for a specific incorrect class. This metric is used to characterize targeted attacks. Experimental results on standard image classification models demonstrate that the two proposed metrics can predict the success rates of FSGM and PGD's targeted or untargeted attacks.

Critical strengths
  1. The writing is easy to follow. The main idea is clear, and the proposed metric is straightforward to understand.

  2. The metrics can quickly predict if model is vulnerable to existing gradient-based attack, e.g., PGD, FSGM. Evaluation is comprehensive on ResNets, ViT, etc.

Critical weaknesses
  1. The related work section is overly lengthy and cites papers mostly before 2020. This suggests that the research direction addressed in the paper has not received much attention recently. Note that, unable to identify an attack on a DNN does not imply that no such attack exists. Quickly identify that existing gradient-based attack have low chance of success doesn't mean other undisclosed gradient-based attacks have low attack success rate.

In a retrospective talk reflecting on 10 years of adversarial machine learning after wrote 9000 adversarial attack papers, Nicholas Carlini noted that attacking DNNs is fundamentally easy. Gradient descent almost always succeeds, whereas establishing effective defenses and rigorous assurance is the actual, challenging problem.

Recent DNN verification provides formal guarantees (e.g., VNN-COMP 2025; Kaulen et al., 2025, "The 6th International Verification of Neural Networks Competition (VNN-COMP 2025): Summary and Results"). Formal verification and formal methods in general is considerably more rigorous than the statistical approach presented in this paper, as it proves unsatisfiability guarantees that no attack exists within the specified perturbation region.

  1. Statement such as "existing robustness evaluations primarily measure overall performance degradation under attack" and "Existing adversarial robustness evaluations primarily rely on aggregated metrics such as accuracy degradation or attack success rate (ASR) to assess overall vulnerability" is not entirely valid.
    State-of-the-art DNN verification rely on certified robustness accuracy and the number of unsatisfiable (UNSAT) local robustness properties (Kaulen et al., 2025), which provide provable guarantees without running attacks. A more interesting question and more related for this paper is determining the tail probability distribution that a class can be attacked. Related issues is actively explored in probabilistic DNN verification, which provides probabilistic bounds on the likelihood of violation (e.g., Hashemi et al., "Scaling Data-Driven Probabilistic Robustness Analysis for Semantic Segmentation Neural Networks"). The authors should explain the merit of this paper in the context of these lines of work.

  2. It is unclear how the predictive power of the proposed metric systematically scales as the number of classes increases or how it performs under regression models.

Questions for authors

I have no questions, as I find the paper's timeliness and merit is limited and do not expect a rebuttal to change my evaluation.

4 comments
AuthorsOpenReview ↗

Response to Weakness 1 & 2 (Scope of SOIV)

We thank the reviewer for emphasizing the importance of recent work on robustness verification and for highlighting the need to clarify the scope of our claims. We agree that our Related Work places too much emphasis on classical attack methods and insufficiently covers formal and probabilistic approaches. We will shorten the discussion of early attacks and incorporate the VNN-COMP 2025 report (Kaulen et al., 2025) and the probabilistic verification approach for semantic segmentation proposed by Hashemi et al. (2025). We will also clarify that our discussion of aggregate accuracy and ASR concerns empirical attack-based evaluation rather than the robustness literature as a whole.

We recognize that low success rates under known attacks do not guarantee robustness against unknown attacks. Formal verification addresses specified robustness properties within a defined perturbation region. For a correctly encoded robustness query, an UNSAT result rules out counterexamples within that region. Probabilistic verification instead provides guarantees for defined safety or violation events under an uncertainty model. We acknowledge the importance of both approaches, but SOIV addresses a different objective. It neither certifies the absence of adversarial examples nor estimates the success probability of an arbitrary attack. The term “attack-agnostic” refers specifically to how SOIV is computed using clean validation predictions, without generating adversarial examples or optimizing against a particular attack. Attacks are used only to evaluate the empirical agreement between SOIV and observed class-wise ASR. This agreement across the evaluated attacks does not imply a guarantee for untested attacks.

Nevertheless, SOIV has distinct practical merits as a lightweight complementary diagnostic tool. First, it can reveal relative vulnerability patterns before performing attacks or verification, using only clean softmax outputs. In the CIFAR-10 and VGG-16 runtime comparison in Section 5.1, SOIV-based analysis required 0.08 minutes versus 13.03 minutes for C&W-based evaluation, approximately 162 times faster in that setting. For a 100-class model, SOIV-T scores for all 9,900 source–target pairs can be obtained from a single set of clean outputs, whereas attack-based evaluation requires a separate targeted attack for each pair. Second, SOIV-U identifies relatively vulnerable classes, while SOIV-T characterizes source-to-target tendencies from the same clean predictions. Section 4.3 combines these views in a single variance and covariance matrix, making both class-wise and directional structures accessible within one framework. Third, these signals can help prioritize classes or class pairs for subsequent attack-based testing, verification, or defense efforts. We do not suggest that verification methods cannot examine individual classes. The benefit is obtaining an initial prioritization signal without first specifying and solving separate verification problems.

The defense application in Section 5.2 provides a concrete example of this utility. Under the same adversarial training budget, allocating adversarial samples according to SOIV-U reduced overall AutoAttack ASR from 61.43% with uniform allocation to 58.17%, while clean accuracy increased from 82.43% to 85.17% (Table 3). Improvements were not uniform across all classes, and prioritization for verification remains a potential application rather than an experimentally established outcome. Thus, we position SOIV as a practical pre-screening tool that can guide where to focus further analysis, not as a substitute for the guarantees offered by formal or probabilistic verification. We will revise the Abstract, Introduction, and Related Work to make these distinct benefits and limitations explicit.

AuthorsOpenReview ↗

Response to Weakness 3 (Scale of SOIV)

We thank the reviewer for raising concerns about scalability with increasing class count and applicability to regression models. We agree that understanding how the predictive value of SOIV changes as the number of classes grows is important. To examine its applicability beyond 10-class classification, we have already evaluated two adversarially trained models on CIFAR-100. Results for PreActResNet-18 and ResNet-18 are reported in Appendix Sections C.4 and C.5, Tables 19 to 22. We also evaluated ImageNet-100 using ViT-L/16, with results in Appendix Section C.6, Tables 23 and 24.

On CIFAR-100, SOIV-U achieves Spearman correlations of 0.8810 and 0.8867 with untargeted MIXED ASR on PreActResNet-18 and ResNet-18, respectively. SOIV-T achieves correlations of 0.8898 and 0.8751 with targeted MIXED ASR on the same models. The targeted experiments compare the vulnerability of 99 potential target classes for a fixed source class, showing that the observed alignment extends beyond class-wise vulnerability to directional relationships in a larger label space. Furthermore, comparisons against uncertainty-related metrics in Appendix Section D.3, Tables 28 and 29, show that SOIV-U maintains relatively high correlations compared with these baselines. On ImageNet-100, the corresponding MIXED Spearman correlations are lower but remain significantly positive, reaching 0.6145 for SOIV-U and 0.6519 for SOIV-T. Taken together, these results provide empirical evidence that SOIV remains informative in multiple 100-class settings, although its predictive strength varies across datasets and models.

To isolate the effect of the number of classes, we subsample K classes from the 100 classes of each adversarially trained CIFAR-100 model. We then recompute the correlation between SOIV-U and MIXED untargeted ASR over only those classes, following the setting of Tables 19 and 21 in the appendix. For K = 10, 20, and 50 we repeat the random subsampling 30 times and report the mean and standard deviation. K = 100 uses all classes, so no repetition is needed. The following tables report the results for the adversarially trained PreActResNet-18 and ResNet-18 respectively.

KPearsonSpearmanKendall
100.8540 ± 0.06660.8295 ± 0.11420.6904 ± 0.1292
200.8481 ± 0.04920.8475 ± 0.05940.6782 ± 0.0748
500.8545 ± 0.01830.8762 ± 0.02080.6962 ± 0.0290
1000.85520.88100.6971
KPearsonSpearmanKendall
100.8672 ± 0.06260.8435 ± 0.07900.6959 ± 0.1003
200.8549 ± 0.05420.8543 ± 0.06820.6868 ± 0.0830
500.8575 ± 0.01910.8795 ± 0.02070.6955 ± 0.0269
1000.86080.88670.6981

Across both models, the mean correlations do not decrease as K increases. The standard deviation decreases, so the correlation becomes more stable with more classes.

Regarding regression models, the current framework is designed for multi-class classification. SOIV-U and SOIV-T rely on class-conditioned softmax probabilities and relationships between source and target classes, which are generally absent from regression outputs. The broader idea of analyzing second-order output statistics may still be relevant, but extending SOIV would require a task-specific definition of adversarial vulnerability, appropriate statistics for continuous predictions, and empirical validation. We will explicitly acknowledge this limitation and identify regression as a direction for future research.

ICLR rule:

There is a word limit for each comment, but there is no limit on the number of comments . You may post multiple replies to each reviewer, but please remain civil and considerate of the reviewers time, as they are reviewing and thus engaging in discussions for multiple papers for ICLR. Compared to other conferences in the field, the open nature of the discussion phases allows more interaction among the authors and reviewers. Please see previous versions of ICLR OpenReview to get a sense of the interactions.

Response is late and long, giving no chance for review to engage in follow up.

I'll keep my score.

We sincerely apologize for our late and lengthy response, which left insufficient time for further discussion. We understand your concerns and respect your decision. Thank you for your time and effort in reviewing our paper and providing valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.