- 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 denote the classifier parameters and the expected cross-entropy under the fixed perturbation distribution , 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 is the loss-induced input density, its normalizing constant, the differential entropy, and 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 is the infimum of the scaled cross-entropy over the classifier’s misclassification region. Both this region and its losses depend on . Consequently, reducing need not increase the complete bound, and Section 4’s treatment of 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 . With denoting the change from the original classifier to the updated classifier,
$\Delta\operatorname{PR}=\Delta B+\Delta G. $
A positive can therefore be outweighed by a negative . 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 , where is the attack energy combining the distance penalty and negative scaled cross-entropy. Under the usual unit-temperature Langevin convention, a drift step size is coupled with Gaussian noise standard deviation , as in the cited projected-Langevin framework of Lamperski (COLT 2021). Algorithm 1 instead uses a step size and noise standard deviation , 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 , 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 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 on handled in the claimed equivalence between maximizing the bound and minimizing the proposed objective? For cross-entropy, a successful misclassification implies , giving the fixed-threshold bound . 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 , 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 , 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
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 . 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 -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 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 are the training samples. In PAT, the are adversarial examples drawn from .
- What the sum runs over. In TERM, is the size of the training set on which the objective is defined, so the weights are normalized over the whole dataset. In PAT, 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 (Theorem 5).
- What the weights mean. In TERM, the weights are a tilt: they define the objective, and 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 (Proposition 4): their role is to correct for the fact that we sample from , whereas PR and its lower bound are defined under .
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 and , 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).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).
Rebuttal (Part 1)
Thank you for your positive assessment and thoughtful feedback on our work. Please find our responses below:
γ∗ depends on θ (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 padv, 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 to achieve clean accuracy comparable to PAT, particularly on ResNet-18. The following results supplement Table 1:
While DA achieves competitive PR at smaller perturbation scales, its performance deteriorates sharply as ϵ increases. At ϵ=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.