acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic Adversarial Training

13 views

Abstract

Building on a probabilistic perspective in which adversarial examples arise from the overlap between a distance-based distribution and a victim-classifier-induced distribution , we start from a simple intuition: adversarial examples become harder to generate when these two distributions are pushed apart, as their overlap becomes smaller, thereby increasing robustness. This intuition naturally motivates a KL-based robustness objective. We then prove that is a lower bound on probabilistic robustness (PR), where denotes the normalizing constant of . Since PR is generally intractable to compute directly, maximizing this KL-based lower bound provides a tractable surrogate objective for improving PR. We further show that this objective recovers a scaled form of adversarial training, offering a probabilistic interpretation of adversarial training and a principled route to robustness improvement. We call the resulting method **probabilistic adversarial training**. Experiments show that it consistently improves PR, and ablation studies demonstrate that the induced scaling factor can even enhance the PR of non-probabilistic adversarial training methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Probabilistic Robustness, Probabilistic Modelling

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_HaK7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Probabilistic Adversarial Training (PAT), which targets classification accuracy under a specified perturbation distribution. It constructs an adversarial distribution by combining a distance-based density with an exponentiated classification loss, generates perturbations using a projected, gradient-clipped Langevin-style procedure, and updates the classifier using detached negative-exponential loss weights. The theoretical analysis provides a KL-form lower bound on probabilistic robustness, an idealized importance-sampling estimator, and a characterization of the practical cross-image estimator through reweighted entropic risk. Experiments on CIFAR-10 and CIFAR-100 with ResNet-18 and WRN-50-2 report improved probabilistic robustness alongside reduced worst-case robustness.

Critical strengths

S1. The paper addresses a meaningful distinction between distributional and worst-case robustness.

S2. Appendix C usefully characterizes the effect of practical batch normalization of the weights.

Critical weaknesses

W1. The theoretical analysis does not establish the claimed improvement in probabilistic robustness or justify the implemented sampling procedure.

Let θ\theta denote the classifier parameters and R(θ)R(\theta) the expected cross-entropy under the fixed perturbation distribution pdisp_{\mathrm{dis}}, for a fixed clean input and true label. Expanding the Kullback–Leibler divergence gives

$D_{\mathrm{KL}}(p_{\mathrm{dis}}|p_{\mathrm{vic},\theta}) -\log Z_{\mathrm{vic}}(\theta) = -H(p_{\mathrm{dis}})-c_2R(\theta), $

where pvic,θp_{\mathrm{vic},\theta} is the loss-induced input density, Zvic(θ)Z_{\mathrm{vic}}(\theta) its normalizing constant, H(pdis)H(p_{\mathrm{dis}}) the differential entropy, and c2>0c_2>0 the fixed loss-scaling coefficient. Thus, maximizing the proposed objective is exactly equivalent to minimizing expected cross-entropy under the perturbation distribution. This identity also appears in the proof of Proposition 4. However, Theorem 2 gives the complete bound

$ \operatorname{PR}(\theta)\ge B(\theta) := 1-\frac{c_2R(\theta)}{\gamma_*(\theta)}, $

where γ∗(θ)\gamma_*(\theta) is the infimum of the scaled cross-entropy over the classifier’s misclassification region. Both this region and its losses depend on θ\theta. Consequently, reducing R(θ)R(\theta) need not increase the complete bound, and Section 4’s treatment of γ∗\gamma_* as a fixed positive scaling factor requires justification.

More fundamentally, even increasing the complete bound does not guarantee an increase in actual PR. Define the nonnegative bound gap as G(θ)=PR⁡(θ)−B(θ)G(\theta)=\operatorname{PR}(\theta)-B(\theta). With Δ\Delta denoting the change from the original classifier to the updated classifier,

$\Delta\operatorname{PR}=\Delta B+\Delta G. $

A positive ΔB\Delta B can therefore be outweighed by a negative ΔG\Delta G. In this particular bound, the gap arises from discarding the loss on correctly classified inputs and replacing the losses on misclassified inputs by their infimum. Increasing confidence on many already-correct inputs can substantially reduce expected cross-entropy while other inputs cross the decision boundary, increasing the probability of misclassification. The analysis establishes neither tightness at the current iterate, control of changes in the bound gap, nor another sufficient condition transferring surrogate improvement to PR improvement. Theorem 5 establishes estimator consistency at fixed parameters, and Appendix C likewise analyzes the practical estimator under fixed parameters and exact sampling. These results do not establish the advertised optimization-level guarantee that PAT is “theoretically guaranteed to improve PR.”

There is also an unresolved discrepancy between the theoretical sampling distribution and the implemented sampler. The importance-sampling analysis assumes samples from the Gibbs density padv∝exp⁡(−Eθ)p_{\mathrm{adv}}\propto\exp(-E_\theta), where EθE_\theta is the attack energy combining the distance penalty and negative scaled cross-entropy. Under the usual unit-temperature Langevin convention, a drift step size η\eta is coupled with Gaussian noise standard deviation 2η\sqrt{2\eta}, as in the cited projected-Langevin framework of Lamperski (COLT 2021). Algorithm 1 instead uses a step size η=0.3\eta=0.3 and noise standard deviation σ=0.001\sigma=0.001, together with gradient clipping and separate projections, without establishing how these choices approximate the stated target density. Consequently, weights derived from the theoretical density have not been shown to correct the sampling bias of the implemented procedure.

These gaps affect the paper’s central theoretical contribution. The authors should establish sufficient conditions connecting the actual update to PR improvement or qualify improvement as an empirical finding, and reconcile the implemented sampler and importance weights with the assumed target distribution.

W2. The incremental contribution over established objectives is insufficiently isolated, and the evaluation leaves the main performance claim difficult to interpret. Holding the generated perturbations fixed and writing their losses as ℓi(θ)\ell_i(\theta), PAT’s outer update satisfies

$ \sum_i \frac{e^{-\beta\ell_i}}{\sum_j e^{-\beta\ell_j}} \nabla_\theta\ell_i = \nabla_\theta \left[ -\frac{1}{\beta} \log\left(\frac{1}{B}\sum_i e^{-\beta\ell_i}\right) \right]. $

This is the negative-tilt instance of Tilted Empirical Risk Minimization by Li, Beirami, Sanjabi, and Smith (ICLR 2021), including its suppression of high-loss samples. The equivalence concerns the fixed-sample outer update, but it substantially narrows the novelty attributable to the weighting mechanism.
The ideal objective also makes direct training on samples from pdisp_{\mathrm{dis}} an essential comparator. Such distribution-dependent perturbation training was studied by Wang, Webb, and Rainforth in Statistically Robust Neural Network Classification (UAI 2021). Although discussed in related work, this baseline is absent from the main comparison. Furthermore, PAT uses 100 inner iterations while the standard adversarial-training baselines use 10, with different perturbation constraints. The evaluation therefore does not isolate the contribution of probabilistic sampling and weighting from the training distribution and additional computation. Appendix E.3 also leaves the PR distribution’s precise parameterization and aggregation protocol unclear. These omissions matter particularly because PR, rather than worst-case robustness, is the paper’s principal claim.

Questions for authors

Q1. How is the dependence of γ∗\gamma_* on θ\theta handled in the claimed equivalence between maximizing the bound and minimizing the proposed objective? For cross-entropy, a successful misclassification implies fθ≥log⁡2f_\theta\geq\log 2, giving the fixed-threshold bound PR⁡≥1−R0(θ)/log⁡2\operatorname{PR}\geq1-R_0(\theta)/\log 2. What additional guarantee does the KL formulation provide beyond this surrogate-risk bound, and which result establishes improvement of actual PR during training?

Q2. What distribution does Algorithm 1 target under the reported step size, noise scale, gradient clipping, and projection sequence? Please reconcile that distribution with the density used to derive the importance weights. A finite-time sampling analysis or a controlled experiment against a tractable reference distribution would materially affect my assessment of whether the probabilistic interpretation applies to the implementation.

Q3. How does PAT compare with direct perturbation training under the same pdisp_{\mathrm{dis}}, and what contribution remains after accounting for the negative-tilt TERM equivalence of its outer update? Please provide comparisons that control training resources and perturbation distributions, together with variability across training seeds. These comparisons are needed to distinguish the proposed sampling mechanism from established loss weighting and additional inner-loop computation.

Q4. Please specify the exact test-time perturbation distribution, the meaning of ϵ\epsilon, boundary handling, and whether PR measures ground-truth accuracy over all test images, accuracy conditional on clean correctness, or agreement with clean predictions. Please also reconcile the claim of consistent weighting gains with Table 2: on CIFAR-10/ResNet-18, PGD-COR reduces PR(0.2) from 77.03% to 75.97%, while on CIFAR-100/ResNet-18, PAT and PAT(WOS) obtain 59.78% and 59.83%, respectively. Which improvements remain reliable after accounting for training and Monte Carlo uncertainty?

5 comments
AuthorsOpenReview ↗

Rebuttal (Part 1)

Thank you for your positive assessment and thoughtful feedback on our work. Please find our responses below:

γ∗\gamma^* depends on θ\theta (W1a, Q1a)

Thank you for pointing this out. Reviewer gwo1 raised a similar concern, so please refer to our responses to gwo1's W1 and Q1. Your formulation of Q1 was also particularly helpful in deriving the constant in our lower bound.

PR Lower Bound vs. Actual PR Improvement (W1b, Q1b)

Thank you for raising this important point. We agree with your observation that increasing even the complete lower bound does not guarantee higher actual PR. Our original claim of a theoretical guarantee was too strong, and we have revised it accordingly, particularly in the third contribution listed in the Introduction.

Nevertheless, optimizing a tractable lower bound is a well-established strategy in machine learning, as exemplified by ELBO optimization and EM algorithm. Similarly, our theory establishes a lower bound on PR and provides a principled surrogate objective for PAT. While this does not guarantee higher actual PR, our empirical results consistently demonstrate improvements across datasets and architectures.

Our revised manuscript therefore distinguishes the theoretical motivation for PAT from the empirically supported improvements in actual PR.

Practical Langevin dynamics and sampling approximation (W1c, Q2)

Thank you for pointing out the discrepancy between the ideal Gibbs distribution and Algorithm 1. We agree with your observation that finite sampling steps, independently chosen step size and noise scale, gradient clipping, and projections introduce sampling bias. Consequently, the resulting finite-step distribution is not guaranteed to equal padvp_{\mathrm{adv}}, and the importance weights do not necessarily correct this discrepancy.

Our implementation is primarily motivated by computational constraints. Adversarial training requires generating adversarial examples for every minibatch, with each inner step involving gradient computation. We therefore adopt a relatively large step size and small noise variance to generate effective energy-guided adversarial examples from uniform noise within 100 steps. Enforcing the standard Langevin relationship would either introduce excessive noise relative to the image range or require substantially more steps. Similar practical approximations have been studied in energy-based modelling [1, 2].

We have clarified this distinction in Section 3 and the newly added Section 5.5, with detailed hyperparameters in Appendix E.2. We do not claim exact sampling or unbiased importance weighting in our implementation; its effectiveness in improving PR is supported by our empirical results.

Comparison with Gaussian Data Augmentation (W2b, Q3a)

Thank you for pointing us to [5]. Under Gaussian perturbations, the training method in [5] amounts to adding Gaussian noise to training inputs, which we refer to as Data Augmentation (DA). We set σ=0.075\sigma=0.075 to achieve clean accuracy comparable to PAT, particularly on ResNet-18. The following results supplement Table 1:

DatasetModelMethodAcc. %PGD %CW %PR0.1PR_{0.1}PR0.12PR_{0.12}PR0.15PR_{0.15}PR0.2PR_{0.2}
CIFAR-10ResNet-18PAT82.4432.5330.5195.5393.4189.2480.62
DA82.180.890.9994.7790.3677.2148.32
WRN-50-2PAT78.2631.9429.9195.9793.9690.9582.19
DA75.540.360.5893.4889.0777.4352.66
CIFAR-100ResNet-18PAT53.5417.5014.8790.7486.5878.0659.74
DA52.890.771.0489.2381.3560.8527.83
WRN-50-2PAT49.9517.2615.9892.2989.4982.1867.15
DA43.860.300.3585.1176.4957.8530.07

While DA achieves competitive PR at smaller perturbation scales, its performance deteriorates sharply as ϵ\epsilon increases. At ϵ=0.2\epsilon=0.2, PAT outperforms DA by approximately 30–37 percentage points across all four settings.

Importantly, PAT achieves both higher clean accuracy and higher PR in every setting, demonstrating that its robustness gains do not come at the expense of clean accuracy. Moreover, DA exhibits near-zero PGD/CW robustness, whereas PAT retains substantial adversarial robustness.

Overall, PAT demonstrates a significantly better accuracy–robustness trade-off, stronger generalization to larger perturbations, and greater adversarial robustness than DA in our experiments.

AuthorsOpenReview ↗

Rebuttal (Part 2)

Comparison with TERM (W2a, Q3b)

We sincerely thank the reviewer for bringing TERM [3] to our attention. We had overlooked this work in our initial literature review, possibly because it does not directly address adversarial robustness. Indeed, TERM explicitly states in Section 5.1 that:

the applicability of TERM to more adversarial forms of robustness would be an interesting direction of future work.

It further clarifies that:

Note that the outliers considered here are unstructured with random noise, and not adversarial.

These statements clearly distinguish the robustness setting considered in TERM from adversarial robustness.

Interestingly, the two methods do arrive at almost the same formula: the softmax-weighted gradient of PAT (Section 5.1, Step 3) has the same form as the tilted gradient of TERM (Lemma 1 in [3]) with t=−c2t=-c_2. TERM itself acknowledges that this form had already been studied in the classical exponential-smoothing literature [4, Proposition 2.1], so it is common to TERM, PAT, and [4]. We believe that sharing a mathematical form does not by itself diminish the novelty of PAT, for two reasons:

(1) Different motivation and derivation. TERM defines the tt-tilted loss directly as an extension of ERM, and the softmax weights appear when the log-sum-exp is differentiated. PAT starts from our KL-based lower bound on PR (Theorem 2), and the softmax weights appear only at the last step, when the gradient of this bound is estimated by SNIS so that the intractable constants ZadvZvicZ_{\mathrm{adv}}Z_{\mathrm{vic}} cancel. The shared form is therefore the common end point of two different derivations, not a shared starting point.

(2) Different usage.

  • What is weighted. In TERM, the xix_i are the training samples. In PAT, the X(i)X^{(i)} are adversarial examples drawn from padv(⋅ ;xori,yori,θ)p_{\mathrm{adv}}(\cdot\,; x_{\mathrm{ori}}, y_{\mathrm{ori}}, \theta).
  • What the sum runs over. In TERM, NN is the size of the training set on which the objective is defined, so the weights are normalized over the whole dataset. In PAT, BB is only the number of Monte Carlo samples in the current batch: normalizing within the batch is the self-normalization step of SNIS, and the estimator is consistent as B→∞B\to\infty (Theorem 5).
  • What the weights mean. In TERM, the weights are a tilt: they define the objective, and tt is a free hyperparameter controlling how strongly high-loss samples are magnified or suppressed. In PAT, the weights are derived as self-normalized density ratios pdis/padvp_{\mathrm{dis}}/p_{\mathrm{adv}} (Proposition 4): their role is to correct for the fact that we sample from padvp_{\mathrm{adv}}, whereas PR and its lower bound are defined under pdisp_{\mathrm{dis}}.

We sincerely appreciate the reviewer for pointing out this relevant work. We have revised Section 7 accordingly.

PR evaluation protocol (Q4a)

Thank you for requesting these clarifications. Our PR evaluation follows the protocol and implementation of AT-PR [6], using the Gaussian perturbation option provided in their code. We acknowledge that these details were insufficiently specified in our original submission.

For each image xi∈[0,1]dx_i\in[0,1]^d and ϵ∈0.1,0.12,0.15,0.2\epsilon\in{0.1,0.12,0.15,0.2}, we generate perturbations as

\tilde{x}_{ij} = \mathrm{clip}_{[0,1]^d}\left(x_i + \mathrm{clip}_{[-\epsilon,\epsilon]^d}(\delta_{ij})\right).$$ Thus, $\epsilon$ is the per-coordinate standard deviation of the Gaussian noise before clipping, and the resulting perturbations satisfy $\\|\tilde{x}_{ij}-x_i\\|_\infty\leq\epsilon$. For each model, we retain the correctly classified images from the 10,000 image test set, denoted by $\mathcal{I}_\theta=\{i:s_i(x_i)<0\}$, where $s_i$ is the logit margin defined in Section 2.3 with respect to the ground-truth label $y_i$. We independently sample $N=100$ perturbations per retained image and compute $$\widehat{\mathrm{PR}}(\epsilon) = \frac{1}{|\mathcal{I}_\theta|N} \sum_{i\in\mathcal{I}_\theta} \sum_{j=1}^{N} \mathbf{1}\\!\left\\{s_i(\tilde{x}_{ij}) < 0\right\\}.$$ We have clarified this evaluation protocol and its aggregation formula in Appendix E.3. ## **About Ablation (Q4b)** Thank you for pointing this out. We agree that the wording in Section 6.2 could be more precise and have revised it accordingly. In particular, PGD-COR is an exploratory extension, and consistent improvements are not guaranteed since PGD does not sample from $p_{\mathrm{adv}}$. We have also clarified that PAT generally, rather than universally, outperforms PAT(WOS).
AuthorsOpenReview ↗

Rebuttal (Part 3)

Reference

[1] Nijkamp, Erik, et al. "Learning non-convergent non-persistent short-run mcmc toward energy-based model." NeurIPS (2019).

[2] Grathwohl, Will, et al. "Your classifier is secretly an energy based model and you should treat it like one." ICLR (2020).

[3] Li, Tian, et al. "Tilted empirical risk minimization." ICLR (2021).

[4] Pee, E. Y., and Johannes O. Royset. "On solving large-scale finite minimax problems using exponential smoothing." Journal of optimization theory and applications 148.2 (2011): 390-421.

[5] Wang, Benjie, Stefan Webb, and Tom Rainforth. "Statistically robust neural network classification." UAI (2021).

[6] Zhang, Yi, et al. "Adversarial training for probabilistic robustness." ICCV (2025).

Reviewer_gwo1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Probabilistic Adversarial Training, a method designed to improve probabilistic robustness. The authors derive a KL-based lower bound for PR and use Langevin dynamics and importance weighting to develop a practical training algorithm. Experiments on CIFAR-10 and CIFAR-100 show improvements in PR over several baselines, although worst-case robustness may decrease.

Critical strengths
  1. The paper provides a clear probabilistic perspective on adversarial robustness and connects adversarial attacks with probabilistic robustness.
  2. The authors provide theoretical analysis and a practical training algorithm.
  3. Experiments on multiple datasets and architectures show promising improvements in PR.
Critical weaknesses
  1. Theoretical limitations of Theorem 2. Theorem 2 provides a lower bound on PR, but it is not fully clear how optimizing the proposed surrogate objective guarantees an improvement in actual PR. In particular, the lower bound also depends on γ∗\gamma^*, which may change with the model parameters. The authors should clarify the conditions under which optimizing this objective leads to better PR.

  2. Fairness of the experimental comparison. The baseline methods use perturbations bounded by L∞≤8/255L_\infty \leq 8/255, while PAT does not impose the same hard perturbation constraint. Therefore, the improvement in PR may partly come from training on a broader or larger range of perturbations, rather than from the proposed importance-weighting mechanism itself. A comparison under matched perturbation distributions would help clarify the source of the improvement.

Questions for authors
  1. Since the lower bound in Theorem 2 depends on γ∗\gamma^*, how can optimizing the proposed surrogate objective reliably improve actual PR? Under what conditions does this guarantee hold?
  2. Can the authors compare PAT with baseline methods under matched perturbation distributions and budgets to determine whether the improvement comes from the proposed weighting mechanism or from differences in training perturbations?
3 comments

Thank you for your positive assessment and thoughtful feedback on our work. Please find our responses below:

Theoretical limitations of Theorem 2. (W1, Q1)

Thank you for pointing out this important issue! You are absolutely right that γ∗\gamma_* depends on θ\theta. Consequently, the inequality in Theorem 2, PR(pdis,s)≥1−log⁡Zvic−H(pdis)−KL(pdis∣pvic)γ∗\mathrm{PR}(p_{\mathrm{dis}}, s) \ge 1 - \frac{\log Z_{\mathrm{vic}} - H(p_{\mathrm{dis}}) - \mathrm{KL}(p_{\mathrm{dis}} \\| p_{\mathrm{vic}})}{\gamma_*} does not, by itself, establish that KL(pdis∣pvic)−log⁡Zvic\mathrm{KL}(p_{\mathrm{dis}} | p_{\mathrm{vic}}) - \log Z_{\mathrm{vic}} is a lower bound on PR(pdis,s)\mathrm{PR}(p_{\mathrm{dis}}, s).

Motivated by reviewer HaK7's Q1, we realized that we can derive a lower bound on PR that is independent of γ∗\gamma_* by establishing a constant lower bound on γ∗\gamma_*.

Let Eadv={x∈X:s(x)≥0}E_{\mathrm{adv}} = \{x \in \mathcal{X} : s(x) \ge 0\} as the vulnerable space where the classifier is fooled. We define γ∗:=inf⁡x∈Eadvc2f(x,yori;θ).\gamma_* := \inf_{x \in E_{\mathrm{adv}}} c_2 f(x, y_{\mathrm{ori}}; \theta). For any x∈Eadvx\in E_\mathrm{adv}, at least one class y≠yoriy\neq y_\mathrm{ori} has a logit no smaller than that of yoriy_\mathrm{ori}, thus pθ(yori∣x)≤1/2p_\theta(y_\mathrm{ori} \mid x)\leq 1/2. As ff is cross-entropy loss, we have f(x,yori;θ)=−log⁡pθ(yori∣x)≥log⁡2f(x, y_{\mathrm{ori}}; \theta) = -\log p_\theta(y_\mathrm{ori} \mid x)\geq \log 2. Hence γ∗=inf⁡x∈Eadvc2f(x,yori;θ)≥c2log⁡2\gamma_{*} = \inf_{x \in E_{\mathrm{adv}}} c_2 f(x, y_{\mathrm{ori}}; \theta) \geq c_2 \log2.

Therefore, we obtain the following new lower bound on PR: PR(pdis,s)≥1−log⁡Zvic−H(pdis)−KL(pdis∣pvic)c2log⁡2\mathrm{PR}(p_{\mathrm{dis}}, s) \ge 1 - \frac{\log Z_{\mathrm{vic}} - H(p_{\mathrm{dis}}) - \mathrm{KL}(p_{\mathrm{dis}} \\| p_{\mathrm{vic}})}{c_2 \log 2} which no longer depends on γ∗\gamma^*.

Based on the above analysis, we have revised the relevant discussions of Theorem 2 in Section 4 and the Appendix B of the revised manuscript.

We sincerely appreciate your constructive feedback, which has helped us strengthen one of the key theoretical results.

Fairness of the experimental comparison. (W2)

Thank you for raising this important point. We agree that PAT is not constrained by the hard L∞L_\infty bounds. However, this difference is an inherent part of PAT's probabilistic formulation: adversarial examples are generated through sampling from padvp_\mathrm{adv}, where pdisp_\mathrm{dis} softly penalizes deviations from the original image. Thus, we are not simply enlarging the perturbation budget, The potentially larger perturbations is from our designed sampling mechanism.

To ensure a fair comparison, we calibrated PAT's parameters so that the resulting classifiers achieve clean accuracy approximately comparable to those trained using the baseline methods. This choice is motivated by the well-established trade-off between clean accuracy and adversarial robustness: increasing the perturbation budget of conventional adversarial training may improve robustness but typically reduces clean accuracy. Since PAT and the baseline methods employ different perturbation mechanisms, directly matching the geometric magnitudes of their training perturbations would not necessarily provide a meaningful comparison. Instead, we compare the PR achieved at similar clean-accuracy levels. Thus, although PAT may generate perturbations with larger geometric deviations, its improvements in PR are achieved while maintaining comparable clean accuracy (Table 1).

We acknowledge that our original submission did not clearly explain that we ensured a fair comparison by approximately matching the clean accuracy across different methods. We have therefore revised Section 6.1 and Appendix E.2 to explicitly clarify this experimental setup. Your feedback has helped us clarify this point considerably.

About Ablation (Q2)

Thank you for pointing this out. In Table 2, we have already conducted an ablation study to evaluate the contribution of the weighting mechanism by comparing PAT with PAT (WOS).

Thank you for the clarifications and revisions, particularly regarding the theoretical analysis of Theorem 2. However, some concerns regarding the fairness of the experimental comparison remain, and I will therefore maintain my original rating.

Thank you again for your positive assessment and constructive comments. We appreciate your feedback, which has helped us clarify several points and improve the manuscript.

Reviewer_m6q8OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a novel probabilistic adversarial training (PAT) method. Specifically, the paper introduces probabilistic robustness (PR), and PAT improves the attack success probability against adversarial distributions. First, the paper extends an existing attack to define an untargeted probabilistic adversarial attack. Then, to address the intractability of PR estimation, the paper first proposes a KL-divergence lower bound as a surrogate objective. This lower bound remains intractable, so the paper further derives a computable training loss via an equivalent importance-sampling-based loss function with normalization that cancels the intractable terms. The paper also provides a theoretical convergence analysis and a practical adjustment to reduce computational cost. The paper compares the proposed method's performance with competitors through experiments.

Critical strengths
  1. The lower-bound and empirical-loss design is theoretically justified. This strengthens the paper’s findings. Although the paper includes many mathematical notions and theorems, the writing is well-structured and easy to follow overall.
  2. The paper compares performance against many competitors for both worst-case robustness and probabilistic robustness. The experiment demonstrated clear improvement over the direct competitor (AT-PR).
Critical weaknesses
  1. The paper should better justify the motivation for improving PR. While the paper clearly demonstrates PR improvements, accuracies on PGD- and CW-generated samples dropped significantly more than the PR improvement. This may be an inevitable trade-off, and the paper briefly discusses that PAT naturally penalizes extreme outliers rather than overfitting. However, the discussion does not make the practical benefit (relative to the degradation in worst-case robustness) very clear.
  2. It would be better to cover more datasets, more model architectures, and more PR defenses. Specifically, the paper evaluates only CIFAR-variant datasets and ResNet-variant architectures, so some practitioners may care about this limited dataset coverage. Also, the related work contains some more PR defenses. The authors justify the disadvantages of those methods, but including comparisons to those defenses would strengthen the experimental support. Given the paper's theoretical depth, the experimental coverage may not be a critical issue.
Questions for authors
  1. If possible, briefly justify the benefit of pursuing PR over worst-case robustness. If the worst-case robustness is just overfitting the extreme samples, is the standard adversarial training losing some advantage? Can this overfitting behavior be experimentally verified to reveal the existing problem of standard adversarial training?
  2. Is there any existing work on the trade-off between worst-case robustness and probabilistic robustness? If this is inevitable, how would practitioners balance those two?
  3. While not necessary, more experiments across datasets and model architectures would improve the experimental demonstration.
  4. In Section 5.4, the paper noted the computational cost of using multiple perturbation samples per image and ultimately used one perturbation per image in the batch. I expect a trade-off between performance and computational cost (using multiple perturbations). Consider additional experiments on the effect of the number of per-image perturbations.
3 comments

Thank you for your positive assessment and thoughtful feedback on our work. Please find our responses below:

Practical Benefit and Trade-off Between PR and WCR (W1, Q1)

Thank you for raising this important question. We have expanded Section 7 to clarify the practical motivation for PR, particularly its relevance to naturally occurring perturbations and quantitative safety assurance [1].

As detailed in our response to Reviewer HaK7 (W2b, Q3a), we additionally compared PAT with Gaussian Data Augmentation (DA) at approximately matched clean accuracy. Together with the baselines in Table 1, these results demonstrate that although PAT is derived from maximizing a lower bound on PR, it retains substantial PGD/CW robustness: PAT improves PR over conventional adversarial training at some cost to PGD/CW accuracy, while outperforming DA on both PR and PGD/CW accuracy.

Thus, PAT's practical benefit lies in improving probabilistic robustness without sacrificing resistance to adversarial attacks altogether.

More Experiments Across Datasets, Model Architectures and PR Defences (W2, Q3)

Thank you for raising this point. The Gaussian data augmentation (DA) discussed above is a representative PR defence. We are also extending our evaluation to an additional dataset and an additional model architecture, and will report the results here before the discussion period ends.

Related Work and Practical Trade-off Between PR and WCR (Q2)

Thank you for raising this important question. We have expanded the Defenses Targeting Probabilistic Robustness paragraph in Section 7 to acknowledge prior studies on the spectrum between random and worst-case robustness [2], alongside the already cited PR-oriented methods.

In practice, the appropriate balance depends on application-specific risk requirements. WCR is particularly relevant when deliberate adversarial attacks are anticipated, whereas PR quantifies reliability under a specified perturbation distribution. Practitioners may therefore prioritize one criterion or seek a balance between both.

As discussed in our response to W1,Q1, PAT provides such a balance by improving PR while retaining substantial adversarial robustness, as supported by our comparisons with conventional adversarial training and Gaussian Data Augmentation (DA).

Number of per-image perturbations (Q4)

Thank you for this suggestion. With one perturbation per image, the per-image weights are identically one, so the K=1K = 1 case is already reported as PAT (WOS) in Table 2.

Going beyond K=1K = 1 is costly for two reasons. First, sample generation dominates the training time and scales linearly with KK: a single CIFAR-10 run takes about 12 hours at K=1K = 1 on one RTX 5090 (Appendix E.4), and therefore about 24 hours at K=2K = 2. Second, with K>1K > 1, the hyperparameters should be re-calibrated to match the clean accuracy of the other methods, as required by our comparison protocol (Section 6.1, Appendix E.2; see also our response to Reviewer gwo1, W2). This calibration itself takes several such 12×K12\times K-hour runs for each value of KK.

We are therefore unable to provide these results before the end of the discussion period, but we will aim to include this ablation in the camera ready version.

Reference

[1] Zhao, Xingyu. "Probabilistic robustness in deep learning: A concise yet comprehensive guide." Adversarial Example Detection and Mitigation Using Machine Learning. Cham: Springer Nature Switzerland, 2026. 209-222.

[2] Rice, Leslie, et al. "Robustness between the worst and average case." NeurIPS (2021)

I appreciate the authors' effort on the rebuttal.

I keep my score and confidence.

Thank you for your response. As promised, we provide the additional results for the new model architecture (VGG19) and dataset (SVHN) below. These results have also been incorporated into Appendix G of our revised manuscript.

New Model Architecture (VGG19)

DatasetModelAttackAcc. %PGD %CW %PR(·, 0.1) %PR(·, 0.12) %PR(·, 0.15) %PR(·, 0.2) %
CIFAR-10VGG19Clean92.860.000.0054.9641.8427.8817.70
FGSM79.9838.5039.1093.2189.9784.3072.66
PGD79.3644.8043.7594.7692.3587.7578.08
ALP79.0443.3942.4195.1192.9388.5979.26
CLP82.4139.3840.2195.3893.1388.6479.35
TRADES77.8245.4443.1595.2693.1089.1381.85
MART76.3648.9643.9295.2393.0989.0581.07
AT-PR78.9243.9342.2195.1892.8788.8179.52
PAT79.1037.9635.2595.6693.6789.9382.90
CIFAR-100Clean72.010.330.1327.7720.2213.106.89
FGSM52.379.2910.0982.0972.1555.2330.95
PGD48.0622.5021.8492.0087.4577.7056.04
ALP49.3821.1420.5592.0387.7578.9460.44
CLP51.1820.9018.8192.1588.2180.4263.82
TRADES48.9223.5220.7090.1984.8075.7957.75
MART45.1423.6220.2291.0486.8078.1258.64
AT-PR47.2421.5420.9592.6588.5879.9162.34
PAT47.6618.6016.1393.3690.3284.1170.74

New Dataset (SVHN)

DatasetModelAttackAcc. %PGD %CW %PR(·, 0.1) %PR(·, 0.12) %PR(·, 0.15) %PR(·, 0.2) %
SVHNResNet-18Clean95.670.490.5193.7191.0086.1176.60
FGSM95.050.660.9395.0092.4287.4276.94
PGD89.4845.1545.8797.4395.5791.0280.26
ALP89.3245.6945.9697.4095.3490.7680.17
CLP90.0450.3949.9497.7696.3992.7882.48
TRADES88.6654.2151.8896.5493.0785.5270.64
MART91.3748.6747.2896.9694.9189.7277.63
AT-PR89.9841.0840.5397.5996.3292.9181.23
PAT90.2814.2514.4798.5898.2197.4094.88
WRN-50-2Clean95.670.350.3993.1090.1384.8275.02
FGSM95.542.762.1897.4095.6390.6974.48
PGD90.7653.8452.7698.2697.3194.6886.47
ALP89.9954.6252.2598.2697.4295.1187.65
CLP92.5450.5651.3797.7196.5593.9086.73
TRADES89.1257.6854.4397.7695.8891.0779.46
MART91.0256.8550.8997.9296.7693.7285.18
AT-PR89.8549.9448.8698.3497.8595.5288.34
PAT91.9216.1216.9498.7698.4397.8196.03

Thank you again for your positive assessment and feedback. We sincerely appreciate your support and thoughtful suggestions, which have helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.