acceptodds
Under review as a conference paper at ICLR 2027

First-Order Characterization of Probabilistic Robustness for Classifiers via the Coarea Formula

Abstract

Adversarial training (AT) is a standard approach for improving worst-case robustness, but its worst-case formulation can be overly conservative when perturbations are predominantly stochastic. This has motivated PR-targeted training, where existing methods typically rely either on surrogate risk objectives or on adaptations of adversarial-search procedures, leaving the first-order behavior of PR itself largely unexplored. In this paper, we study the first-order optimization structure of PR and ask when its variation can be represented by only a few perturbations. Under mild regularity conditions, we express the PR gradient via the co-area formula as a distribution-weighted integral over the decision boundary, revealing an inherently aggregate boundary effect. To study this quantity in high dimensions, we use a Gaussian kernel as a computational probe, yielding a simple Monte Carlo estimator of the PR first variation. Low- and high-dimensional experiments show that the extent to which a small number of perturbations represent the aggregate PR variation depends strongly on the underlying geometry and dataset. We further study how the estimated PR variation depends on the chosen margin level and smoothing bandwidth, and explore its use in PR-oriented training.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Probabilistic Robustness, Adversarial Training

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HDZ6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies the first-order behavior of Probabilistic Robustness (PR) with respect to model parameters, provide a new perspective for understanding the first-order optimization structure of PR. By applying the co-area formula, the authors formulate the gradient of the Non-Robust Probability (NPR) as a weighted integral over the decision boundary and approximate it using Gaussian kernels and Monte Carlo sampling. Based on this gradient estimator, they develop NPR-Aug, a training strategy designed to improve probabilistic robustness. Experiments show that aggregating multiple perturbations yields more reliable NPR descent directions and enhances robustness against stochastic perturbations.

Critical strengths
  1. The paper derives a boundary-integral representation of the NPR gradient using the co-area formula, providing a principled characterization of how decision-boundary contributions determine the first-order variation of PR.
  2. The experiments support the theoretical analysis by demonstrating the importance of aggregating boundary contributions. The proposed NPR-Aug. further shows that incorporating this information into training can improve probabilistic robustness.
Critical weaknesses
  1. The high-dimensional experiments are primarily conducted on CIFAR-10 with ResNet-18, without evaluations across diverse datasets or model architectures. As such, the current experimental scope is inadequate to demonstrate the generality of the proposed findings.
  2. Systematic ablations on perturbation sample size, gradient aggregation strategies, and NPR loss weight are missing, making it difficult to assess the contributions of these design choices.
Questions for authors
  1. Could the authors extend the experiments in Section 6.3 to a more challenging dataset, such as CIFAR-100 or ImageNet, and a different model architecture, such as Vision Transformer?
  2. A sensitivity analysis of NPR-Aug. with different perturbation sample sizes (N=1,2,4,8N=1,2,4,8) and NPR loss weights (λ\lambda) under identical training settings would help clarify the impact of these hyperparameters on performance.
2 comments
AuthorsOpenReview ↗

W1/Q1. Additional datasets and architectures

Thank you for recognizing the theoretical characterization and the value of aggregating boundary contributions. We appreciate your suggestions for broader evaluation and systematic ablations, which help clarify the role of each design choice.

To extend Section 6.3 beyond CIFAR-10/ResNet-18, we added CIFAR-100 and VGG16-bn experiments. PR below averages uniform, clipped Gaussian, clipped Laplace, sign, sparse and low-frequency noise, with identical evaluation perturbations within each setting. Accuracies are percentages. NPR-Aug. uses 4×4\times augmentation, where 1×=8/2551\times=8/255.

1. CIFAR-100 with ResNet-18

The standard-training reference has 77.58 clean accuracy and PR of 29.25, 15.52 and 6.00 at 4×4\times, 8×8\times and 16×16\times. The table compares robustness-training methods. MMA and NPR-Aug. report mean ± sample standard deviation over three seeds. Other rows are single runs.

MethodCleanPR at 4×PR at 8×PR at 16×
PGD-AT58.2649.1425.589.58
TRADES52.7846.6728.4111.81
AT-PR57.2850.1726.7510.35
EVaR75.5842.1219.688.11
MMA60.67 ± 0.4754.57 ± 0.5434.95 ± 0.3313.37 ± 0.25
NPR-Aug.63.05 ± 0.1756.04 ± 0.2334.28 ± 0.6113.26 ± 0.29

MMA and NPR-Aug. achieve the highest mean PR among the evaluated methods at all three radii. Both exceed every other method by at least 4.4 points at 4×4\times and 5.8 at 8×8\times. EVaR has the highest clean accuracy among the defended models but lower PR under these stronger perturbations.

NPR-Aug. further improves clean accuracy and PR at 4×4\times over MMA. The gains are 2.38 points (95% CI [1.35, 3.41]) and 1.46 points ([0.32, 2.61]), respectively. At larger radii, NPR-minus-MMA differences are small and their confidence intervals include zero.

A geometric interpretation. Anchor cosine measures alignment between the anchor gradient and a sampled aggregate of boundary contributions. On NPR-Aug. checkpoints, anchor cosine is 0.18 higher on CIFAR-100 than on CIFAR-10 (three seeds per dataset, 95% CI [0.09, 0.28]). For MMA, the difference is 0.05 with an interval of [−0.01, 0.11]. These observations are consistent with the possibility that a more representative anchor leaves less room for supplementation.

2. VGG16-bn on CIFAR-10

To test a different architecture, we use VGG16-bn, which has no residual connections, with the same training recipe and no hyperparameter retuning. Each row below is one training run. Standard training gives 94.36 clean accuracy and PR of 58.12, 38.96 and 30.07 at the three radii.

MethodCleanPR at 4×PR at 8×PR at 16×
PGD-AT81.8277.7762.3833.55
EVaR93.5670.5440.3427.27
MMA80.7073.9863.0043.66
NPR-Aug.85.0680.8268.3743.33

NPR-Aug. has the highest observed PR at 4×4\times and 8×8\times among these methods. It improves on MMA by 6.84 and 5.37 points and on PGD-AT by 3.05 and 5.99, while also exceeding both on clean accuracy. MMA and NPR-Aug. have similar observed PR at 16×16\times. EVaR retains higher clean accuracy but lower PR.

AuthorsOpenReview ↗

W2/Q2. Ablations on sampling, aggregation and loss weight

We added CIFAR-10/ResNet-18 experiments with the training schedule, optimizer and MMA hinge fixed. PR averages uniform, clipped Gaussian, clipped Laplace, sign, sparse and low-frequency noise, using identical evaluation images and perturbation draws within each comparison. The NN and λ\lambda scans use 4×4\times augmentation and one training run per setting, where 1×=8/2551\times=8/255.

  • How sensitive is NPR-Aug. to the number of perturbations? We vary N∈{1,2,4,8}N\in\{1,2,4,8\} with λ=0.5\lambda=0.5. Time is averaged over recorded training epochs.

    Samples per image NNPR at 4×PR at 8×PR at 16×Seconds per recorded epoch
    183.1571.4042.31138.9
    282.1072.6245.97145.3
    483.7872.0543.13165.6
    882.9271.3140.93190.7

    Increasing NN from four to eight adds about 15% to the recorded epoch time without improving PR at the displayed radii. The best setting varies with evaluation radius, while the paper's N=4N=4 offers competitive performance at moderate sampling cost in this scan.

  • How sensitive is the result to the NPR loss weight? We vary λ\lambda with N=4N=4. The boundary-anchor term becomes (1−λ)Lanchor+λLNPR(1-\lambda)L_{\mathrm{anchor}}+\lambda L_{\mathrm{NPR}}, with other MMA terms unchanged. Thus, λ=0\lambda=0 recovers MMA.

    NPR weight λ\lambdaPR at 4×PR at 8×PR at 16×
    0.0077.0965.7241.05
    0.2581.2468.7241.66
    0.5083.7872.0543.13
    0.7582.3565.4634.83

    The paper's choice of λ=0.5\lambda=0.5 gives the highest observed PR at all three radii in this scan. Increasing it to 0.75 lowers PR at 16×16\times by 8.30 points, supporting a balance between the anchor and sampled terms.

  • How do individual directions compare with aggregated boundary contributions? We vary the number of sampled boundary points KK on three CIFAR-10 MMA training seeds. These are diagnostic boundary points, distinct from the NN random perturbations used in training. Values are mean ± sample standard deviation across runs, after averaging over retained images within each run.

    Boundary points KKSampled-point cosineAnchor cosine
    20.876 ± 0.0080.688 ± 0.015
    40.907 ± 0.0050.688 ± 0.021
    80.920 ± 0.0020.701 ± 0.026
    160.924 ± 0.0040.711 ± 0.029

    Sampled-point cosine compares each point's parameter gradient with the weighted aggregate of the other K−1K-1 points. Anchor cosine compares the MMA anchor's gradient with all KK points. Weights are 1/∥∇δm∥21/\|\nabla_\delta m\|_2, and higher cosine means closer directional agreement. The anchor remains less aligned with the sampled aggregate than the sampled boundary points across all tested reference sizes and seeds. This is a directional diagnostic, not a direct comparison of aggregation strategies during training. Sampled points and retained images also vary with KK.

We will include these experiments and their settings in the revision.

Reviewer_noprOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the first-order variation of probabilistic robustness under model parameter updates, deriving a decision-boundary integral characterization using the co-area formula. It proposes a Gaussian-kernel Monte Carlo estimator of the PR gradient and investigates how well individual perturbations approximate the aggregate boundary contribution.

Critical strengths
  • The paper provides an interesting and mathematically sound characterization of probabilistic robustness, with a clear geometric interpretation of how decision-boundary motion affects misclassification probability.

  • The low-dimensional experiments validate the theoretical characterization well by comparing predicted NPR changes against numerical measurements of decision-boundary motion.

  • The analysis of how individual perturbations represent the aggregate NPR gradient is interesting and provides useful insights into the limitations of single-point boundary-based optimization.

  • NPR-Aug. achieves substantial improvements in probabilistic robustness over several established baselines across multiple perturbation distributions, demonstrating the practical potential of the approach.

Critical weaknesses
  • The paper suffers from limited theoretical novelty and insufficient positioning against prior work. The main derivations are mainly applications of calculus theory that has been around for decades to the domain of PR.The main theoretical result, expressing the first variation of NPR as a probability-weighted integral over the decision boundary (Theorem 4.5 and Corollary 4.6), appears to follow relatively directly from established shape-derivative and moving-domain integration theory. The literature seems to be touching very vlose subjects, yet discussion of relevant prior work is lacking. From a short literature review, Uryasev (1994) studies derivatives of parameter-dependent probabilities defined by inequalities and derives representations involving surface integrals. Earlier robustness studies, e.g. Fawzi et al. (NeurIPS 2016) and Franceschi et al. (AISTATS 2018), show connections between stochastic robustness and decision-boundary geometry. Although these works do not derive the same parameter-space NPR gradient, the manuscript does not sufficiently clarify what new theoretical insight follows from its application of classical calculus beyond the expected observation that changes in adversarial probability mass depend on aggregate boundary motion.

  • The proposed gradient estimator is equivalent to optimizing a conventional smoothed classification objective (The authors do acknowledge in Appendix C) - the proposed estimator is the negative gradient of a probit-smoothed zero-one classification loss, obtained by applying the Gaussian cumulative distribution function to the negative classification margin. Algorithm 1 optimizes this smooth loss. Therefore, the proposed optimization procedure can be obtained through a straightforward Gaussian-CDF approximation to the classification indicator, without requiring the co-area derivation. The geometric interpretation provides a useful motivation for boundary-localized optimization, and the reported robustness improvements are convincing, however the manuscript does not show if the specific proposed formulation provides an advantage over simpler stochastic regularization approaches, such as random-noise cross-entropy or alternative smooth margin-based losses.

  • There are some missing simple baselines and ablations that are needed to explain the reported robustness improvements. NPR-Aug. substantially improves probabilistic robustness over MMA and the other reported methods, but the experiments do not establish whether these gains arise from the proposed boundary-localized gradient or from introducing stochastic regularization into MMA. Furthermore, Algorithm 1 simultaneously introduces the NPR regularizer and reduces the contribution of MMA's original boundary-anchor loss through the coefficient λ\lambda. The effects of these changes are not isolated. Comparisons with matched perturbation sampling, regularization weights, and computational budgets are needed to establish the specific benefit of the proposed formulation.

  • A main claimed contribution is understanding when a small number of perturbations can represent the aggregate NPR variation. However, the experiments mainly demonstrate that individual boundary-based directions can be poorly aligned with the aggregate direction, whereas averaging several sampled contributions gives better alignment. The high-dimensional analysis in Figure 2(a) relies on one CIFAR-10 model checkpoint and uses a 256-sample smoothed-gradient estimate as its reference, with additional kernel-free finite-difference validation. While this provides useful empirical evidence, the paper does not fully characterize the geometric conditions under which individual perturbations become representative, or distinguish the benefits of boundary localization from those of averaging multiple gradient contributions.

  • The robustness comparisons do not fully control for differences in training objectives and perturbation settings. Although the methods share a common training protocol, they optimize substantially different objectives, and some relevant training configurations are not matched. For example, the EVaR baseline is trained at a 1×1\times perturbation scale, whereas the proposed NPR component uses 4×4\times perturbations. Moreover, the improved stochastic robustness comes with a reduction in worst-case adversarial accuracy. Relative to MMA, NPR-Aug. improves uniform-noise PR at R=8×R=8\times from 66.39% to 76.02%, while PGD-10 accuracy decreases from 51.72% to 45.55% and training time increases. This trade-off is reported but not sufficiently analyzed.

  • Statistical reliability is lacking. The entire high-dimensional evaluation is restricted to CIFAR-10 with a ResNet-18. This is a limited basis for general conclusions about neural-network decision-boundary geometry, particularly when the paper itself emphasizes that the representativeness of individual perturbations depends strongly on geometry and dataset. Evaluation on additional datasets, such as CIFAR-100, Tiny-ImageNet, or ImageNet, and different architectures would substantially strengthen the findings. Such broader evaluation is also present in related work such as AT-PR (ICCV 2025). Additionally, Appendix F states that all main robustness results use a single training seed, and several ablations also rely on individual seeds. The absence of repeated-run statistics makes it difficult to assess the reliability of the reported improvements.

Questions for authors
  1. What distinguishes the theoretical characterization from established shape-derivative and probability-gradient formulations, beyond its application to PR?

  2. Did the authors compare NPR-Aug. against MMA augmented with simpler stochastic regularizers, such as random-noise cross-entropy or alternative smooth margin losses, under matched training conditions? This would help establish whether the improvements arise specifically from the proposed formulation.

  3. Can the authors provide additional evidence that the findings generalize across datasets and architectures, particularly given the claimed dependence on decision-boundary geometry?

  4. How consistent are the reported robustness improvements across independent training seeds, given that the main comparison uses a single seed?

7 comments
AuthorsOpenReview ↗

W1/Q1: Theoretical Contribution and Connections to Prior Work

Thank you for the detailed review and for pointing us to relevant prior work.

1. Mathematical foundations and theoretical motivation

Our derivation builds on the classical coarea formula, stated in Theorem 3.6. Under the stated regularity assumptions, Theorem 4.5 and Corollary 4.6 characterize the first variation of non-robustness probability, NPR=1−PR\mathrm{NPR}=1-\mathrm{PR}, with respect to the classifier's parameters. The results cover smooth classifiers, with an extension to piecewise-smooth classifiers under additional conditions in Remark 4.3 and Appendix B.

Uryasev (1994) develops a general mathematical framework for differentiating probabilities defined by inequalities, including surface-integral representations. Our boundary-integral formulation is consistent with this framework, and we will make that connection explicit. We claim no novelty in the underlying calculus.

Our motivation concerns the optimization of PR itself. As discussed in Section 2.2, existing approaches optimize statistics of surrogate losses, such as CVaR (Robey et al., 2022) and EVaR (Zhang et al., 2024a), or adapt adversarial training to PR, as in AT-PR (Zhang et al., 2025a). These approaches provide practical training objectives, but their update directions do not by themselves characterize the first-order change in indicator-defined PR. We therefore ask how PR itself changes under a parameter update. This leads us to examine which perturbations provide representative update directions and when an individual adversarial example is sufficient.

Theorem 4.5 and Corollary 4.6 express this change through a boundary integral combining q(δ)/∥∇δm(δ)∥q(\delta)/\|\nabla_\delta m(\delta)\| with ∇θm(δ)\nabla_\theta m(\delta). The input-gradient norm determines how a margin change translates into boundary displacement, the density weights its effect on probability mass, and the parameter gradient determines how an update changes the margin. Consequently, proximity to the boundary alone does not determine representativeness. A selected point must also reflect the directions and relative weights of contributions across the boundary. This motivates NPR-Aug., which supplements MMA's single-anchor update with a Gaussian approximation to the aggregate PR variation. Under the stated conditions, the estimator's expectation converges to ∇θPR=−∇θNPR\nabla_\theta\mathrm{PR}=-\nabla_\theta\mathrm{NPR} as τ→0\tau\to0, connecting the augmentation to the exact first-order change in PR.

2. Broader related work

We will broaden our related work from PR evaluation and training to robustness under random perturbations in deep learning, including probability derivatives, geometric analyses and randomized smoothing, while distinguishing their objectives. The related work mentioned by reviewer will also be included in our revised version as follows:

Fawzi et al. (2016) study the minimum perturbation norm needed to change a prediction within a randomly chosen subspace. Their bounds connect adversarial and random-noise robustness for affine classifiers and extend to nonlinear classifiers under curvature conditions.

Franceschi et al. (2018) relate boundary distance to the scale of uniform ℓp\ell_p or Gaussian noise needed to change a prediction with a specified probability. Their analysis covers linear classifiers and nonlinear classifiers with locally approximately flat boundaries.

These studies characterize robustness to perturbations of a fixed classifier. Our question is how indicator-defined PR changes as the classifier's parameters are updated. The parameter-space gradient connects input-space geometry to the effect of training updates on probability mass.

What goes beyond applying classical calculus to PR? Our contribution connects the exact first variation of PR to the choice of perturbations used in training. It provides a reference for examining whether an update based on a selected point represents the aggregate PR variation, and motivates supplementing that update when it does not. NPR-Aug. puts this perspective into practice by augmenting MMA’s single-anchor update with an approximation to the aggregate variation. Our directional analyses examine representativeness, while the training experiments evaluate the practical benefit of this augmentation. We will revise the contribution statement to make this connection between theory, perturbation selection and training explicit.

AuthorsOpenReview ↗

W2/Q2: Connection to Exact PR Variation and Comparison with Stochastic Regularizers

1. The role of the coarea derivation

The coarea derivation shows how indicator-defined PR changes under parameter updates, offering a different perspective from prior approaches to PR training. For example, AT-PR aims to improve PR by selecting adversarial examples associated with larger adversarial regions, while still using PGD to find high-loss interior candidates. Our first variation (Corollary 4.6) identifies weighted boundary contributions as governing first-order PR change.

This motivates examining whether MMA's boundary anchor represents the aggregate direction and how to supplement its update when it does not. Our Gaussian estimator approximates this aggregate, with its expectation converging to the PR gradient as the bandwidth vanishes under the stated conditions (Eqs. 20–21). The same smooth objective can be constructed through a Gaussian-CDF relaxation, as acknowledged in Appendix C. The coarea derivation explains its connection to exact PR variation and guides the study of representative perturbations and augmentation.

2. Comparisons with simpler stochastic regularizers

Thank you for recognizing the geometric motivation. On CIFAR-10/ResNet-18, NPR and noise-CE share the MMA base, training schedule, uniform sampling, N=4N=4, λ=0.5\lambda=0.5 and sampled-forward budget. The MMA hinge is 32/25532/255, 1×=8/2551\times=8/255, and NPR uses τ=0.5\tau=0.5. Evaluation uses identical draws on 5,000 images. Each configuration uses at least three independent seeds. Values are percentages (± seed SD). Brackets give pointwise Welch-type 95% confidence intervals without multiplicity adjustment.

Overall comparison. PR is averaged over uniform, clipped Gaussian, clipped Laplace, sign, sparse and low-frequency noise.

Augmentation radiusSampled termPR at 1×PR at 4×PR at 8×PR at 16×
4×NPR88.78 ± 0.3683.20 ± 1.0072.88 ± 0.8244.42 ± 1.25
4×Noise-CE88.92 ± 0.4084.89 ± 0.5972.41 ± 1.8242.08 ± 1.81
16×NPR88.72 ± 0.3682.20 ± 0.8971.94 ± 1.6944.25 ± 3.96
16×Noise-CE87.64 ± 0.3581.10 ± 0.4371.84 ± 1.2942.00 ± 2.93

At 4× augmentation, noise-CE has higher mean PR at 1×/4× evaluation and NPR at 8×/16×. At 16× augmentation, NPR has higher means at all four radii. We conjecture that smaller perturbation regions probe less of the boundary, limiting the benefit of approximating its aggregate contribution.

Selected noise distributions. Under 16× augmentation, entries give NPR minus noise-CE in PR points for uniform (training), low-frequency and sparse noise.

Evaluation noiseDifference at 4×At 8×At 16×
Uniform+1.17 [+0.05, +2.29]+0.38 [−1.79, +2.55]+0.24 [−4.77, +5.25]
Low-frequency+0.98 [+0.48, +1.48]+1.23 [+0.00, +2.45]+0.62 [−1.75, +2.98]
Sparse+1.17 [+0.68, +1.65]+0.95 [+0.16, +1.73]+3.91 [+1.44, +6.38]

NPR has higher means on sparse and low-frequency noise. We next test broader corruptions.

Evaluation on more general corruptions. We evaluate the same 16×-augmented models on 10,000 CIFAR-10-C-Full images per condition. Each row averages 19 corruptions at one severity, and mCA averages all 95 conditions.

SeverityNoise-CE accuracyNPR accuracyNPR minus noise-CE
185.90 ± 0.4387.08 ± 0.39+1.18 [+0.58, +1.78]
282.59 ± 0.5883.97 ± 0.62+1.38 [+0.50, +2.26]
379.09 ± 0.7380.59 ± 0.79+1.50 [+0.39, +2.61]
474.99 ± 0.7476.53 ± 0.86+1.54 [+0.36, +2.71]
569.66 ± 0.8971.17 ± 0.96+1.51 [+0.16, +2.85]
All (mCA)78.45 ± 0.6779.87 ± 0.70+1.42 [+0.42, +2.42]

NPR improves mean accuracy at every severity, with all pointwise intervals above zero. It also has higher means on all 19 corruptions after averaging severities. Its higher clean accuracy (89.26 versus 88.21) may contribute to this gain. Logistic smoothing is also competitive, with rankings varying across settings. Full noise-distribution and logistic comparisons will appear in the appendix.

We indeed found that noise-CE also provides useful supplementation to MMA for PR optimization. However, our proposed method still shows an advantage on the broader set of corruptions tested under 16× augmentation. More fundamentally, our central contribution is to provide a distinct perspective on PR optimization. Motivated by this perspective, we examine whether a single boundary anchor is sufficient to represent the aggregate direction and how to supplement its update when needed. The estimator in Section 5 provides a practical approximation for this purpose. Its theoretical connection to PR variation motivates our formulation, while the experiments identify settings where it offers an advantage over simpler alternatives.

AuthorsOpenReview ↗

W3: Ablations on the Sampled Objective and Anchor Weight

All experiments use CIFAR-10/ResNet-18. Unless stated otherwise, each configuration uses at least three independent training seeds. Accuracy values are percentages, ± denotes sample standard deviations across seeds, and brackets give pointwise Welch 95% confidence intervals.

1. Examining the sampled objective without MMA

To compare the sampled objectives without MMA, we use clean CE plus sampled NPR against clean CE plus noise-CE. Both train on uniform noise. Training and evaluation use 16×, where 1× = 8/255.

Evaluation noiseNoise-CE PRNPR PRRelative transfer advantage Δ\Delta
Uniform79.04 ± 0.6672.86 ± 2.490.00
Clipped Laplace70.08 ± 0.9767.05 ± 2.18+3.16 ± 1.16
Clipped Gaussian71.76 ± 0.6467.87 ± 2.51+2.29 ± 1.08
Low-frequency46.02 ± 0.5047.91 ± 2.14+8.07 ± 3.22
Sign35.47 ± 3.0140.12 ± 1.64+10.83 ± 4.62
Sparse69.03 ± 2.3677.71 ± 1.55+14.87 ± 4.48

For distribution ss, the last column compares each method's PR change from its own uniform-noise reference,

Δ(s)=[PRNPR(s)−PRNPR(uniform)]−[PRCE(s)−PRCE(uniform)]\Delta(s)=[PR_{NPR}(s) - PR_{NPR}(\text{uniform})] - [PR_{CE}(s) - PR_{CE}(\text{uniform})]

We compute this difference within each seed before reporting its mean and standard deviation. The uniform entry is zero by definition. A positive value indicates a more favorable change in PR for NPR when the noise distribution changes. For example, switching to low-frequency noise reduces PR by 33.02 points for noise-CE and 24.95 for NPR, giving Δ=8.07\Delta=8.07.

NPR has a positive relative transfer advantage on all five alternative distributions in every seed. The gains range from 2.29 to 14.87 points. NPR also has higher absolute mean PR on low-frequency, sign and sparse noise. Noise-CE remains ahead on uniform, clipped Laplace and clipped Gaussian noise. The relative-transfer advantage therefore complements the raw PR comparison, whose uniform starting points differ. These results show that the choice of sampled objective affects noise transfer even without MMA, although they do not by themselves isolate boundary localization as the cause. We next test whether adding the sampled term contributes beyond changing the MMA anchor weight.

2. Separating the effects of λ\lambda

NPR-Aug. replaces the boundary-anchor loss with

(1−λ)Lanchor+λLNPR.(1-\lambda)L_{\mathrm{anchor}}+\lambda L_{\mathrm{NPR}}.

To separate these changes, anchor-only keeps (1−λ)Lanchor(1-\lambda)L_{\mathrm{anchor}} and omits the NPR term, leaving the other MMA terms unchanged. This provides a common control for both augmentation radii below.

All modified objectives use λ=0.5\lambda=0.5 and share the MMA hinge and training schedule. NPR-Aug. uses N=4N=4 uniform samples. Models are evaluated on identical images and perturbation draws. Anchor-only omits the sampled forward pass and is cheaper. PR averages uniform, clipped Gaussian, clipped Laplace, sign, sparse and low-frequency noise. The final column gives the PR gain over anchor-only at 4× evaluation, with its 95% confidence interval.

ObjectiveAugmentation radiusClean accuracyPR at 4×PR at 16×Gain over anchor-only at 4×
MMANot used84.37 ± 0.9776.41 ± 0.6040.85 ± 0.21+0.72 [−1.13, +2.57]
Anchor-onlyNot used83.95 ± 1.3575.68 ± 0.9141.42 ± 2.510 (reference)
NPR-Aug.4×89.11 ± 0.4283.20 ± 1.0044.42 ± 1.25+7.51 [+5.35, +9.68]
NPR-Aug.16×89.08 ± 0.4282.20 ± 0.8944.25 ± 3.96+6.51 [+4.73, +8.29]

At the same reduced anchor weight, adding NPR improves PR at 4× evaluation by 7.51 points with 4× augmentation and 6.51 points with 16× augmentation. Both confidence intervals are above zero. Thus, adding the NPR term provides a benefit beyond reducing the anchor weight alone.

Sensitivity to the augmentation weight.

As a supplementary ablation, we vary λ\lambda with N=4N=4 and a 4×4\times augmentation radius, using one training run per setting.

λ\lambdaPR at 4×PR at 8×PR at 16×
0.0077.0965.7241.05
0.2581.2468.7241.66
0.5083.7872.0543.13
0.7582.3565.4634.83

Both λ=0.25\lambda=0.25 and λ=0.5\lambda=0.5 improve PR over λ=0\lambda=0 at all three evaluation radii, further supporting the benefit of augmentation. The paper's choice of λ=0.5\lambda=0.5 gives the highest observed PR throughout this scan. Increasing it to 0.750.75 lowers PR at 16×16\times by 8.30 points, supporting a balance between the anchor and NPR terms rather than simply increasing the augmentation weight.

Reviewer_Z39pOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the parameter derivative side of indicator-defined probabilistic robustness. For a fixed perturbation distribution, it expresses the gradient of non-robustness probability (NPR) as an integral over the classifier's decision boundary, weighted by perturbation density, inverse input-margin gradient norm, and parameter-margin sensitivity. A Gaussian kernel approximates the boundary-supported derivative using ambient perturbation samples; this estimator is equivalently the gradient of an expected smoothed margin loss.

Experiments on two-dimensional Penguins data and CIFAR-10 examine full-boundary versus pointwise directions, kernel location, and bandwidth. An MMA-based training variant combines the probit loss with the boundary-anchor loss and improves stochastic-perturbation accuracy on the reported CIFAR-10 comparison.

Critical strengths
  1. The geometric interpretation is clear and useful. Equations 13–15 distinguish boundary location, probability weighting, and parameter sensitivity. The resulting explanation of why a nearest-boundary point can differ from the NPR gradient is more informative than treating every boundary point as equally representative.

  2. The empirical investigation contains relevant diagnostics and promising results. Figure 1 directly compares boundary predictions with changes in NPR across multiple smooth models and seeds. Figure 2(a) evaluates candidate directions through an indicator-based finite difference, which is a useful attempt to avoid evaluating only the smoothing objective. In Table 1, NPR-Aug. improves uniform-noise PR at R=8×R=8\times from 66.39% to 76.02% relative to MMA, while also improving clean accuracy. The paper reports its accompanying reduction in PGD-10 accuracy and increase in training time.

Critical weaknesses

Due to lack of familiarity to this topic, the reviewer does not identify any critical weakness of this paper.

Questions for authors

Can the paper's existing analyses identify a measurable property of boundary contributions that explains when a few perturbations are representative? For example, do contributions from different boundary locations point in similar parameter-space directions? Is most of the contribution concentrated in a small boundary region? Do large contributions cancel one another? Are different type of perturbations contribute to similar or different parameter-space directions? It would be interesting to see some insights on why and when of the purposed methods will work, in addition to the results presented in this paper (about whether and how it works).

1 comment
AuthorsOpenReview ↗

Response to Reviewer Z39P: Boundary Geometry and Perturbation Representativeness

Thank you for recognizing the geometric interpretation and indicator-based validation. We added experiments to address your questions about why and when individual perturbations are representative.

Our initial diagnostic covers 21 checkpoints across CIFAR-10/100 and ResNet-18/VGG16-bn. We sample eight boundary points per retained image within the valid-pixel 4×4\times budget, where 1×=8/2551\times=8/255. Under uniform noise, their weighted parameter gradients are vk=∇θm(x+δk,y)/∥∇δm(x+δk,y)∥2v_k=\nabla_\theta m(x+\delta_k,y)/\|\nabla_\delta m(x+\delta_k,y)\|_2. We use G=∑kvkG=\sum_kv_k as a sampled reference, not the exact boundary integral. Full sampling and numerical details will be provided in the appendix.

  • Do different boundary locations contribute similar directions? Each row below summarizes three ResNet-18 training seeds as mean ± sample SD. Sampled-point cosine compares each gradient with the aggregate of the other seven. Anchor cosine compares the MMA anchor's parameter gradient cac_a with all eight,

    \qquad A(x)=\cos(c_a,G).$$ Cosines are averaged over retained images within each run. Higher values mean closer agreement. The last two columns are defined below. We report the repeated ResNet-18 results below. The full 21-checkpoint results, including five VGG16-bn checkpoints, will be included in the appendix. | Dataset | Model | Sampled-point cosine | Anchor cosine | Effective points (max. 8) | Coherence | | --- | --- | --- | --- | --- | --- | | CIFAR-10 | MMA | 0.920 ± 0.002 | 0.701 ± 0.026 | 7.941 ± 0.012 | 0.940 ± 0.002 | | CIFAR-10 | NPR-Aug. | 0.912 ± 0.002 | 0.545 ± 0.042 | 7.944 ± 0.003 | 0.933 ± 0.002 | | CIFAR-100 | MMA | 0.909 ± 0.006 | 0.748 ± 0.006 | 7.964 ± 0.002 | 0.931 ± 0.004 | | CIFAR-100 | NPR-Aug. | 0.904 ± 0.004 | 0.728 ± 0.014 | 7.961 ± 0.003 | 0.927 ± 0.003 | Mean sampled-point cosine is 0.904–0.920, compared with anchor cosine of 0.545–0.748. *Sampled-point alignment exceeds anchor alignment in the initial 21 checkpoints and all 12 runs summarized here.* Thus, sampled contributions can agree closely even when the selected anchor is less representative.
  • Is most of the contribution concentrated in a small region? We measure concentration among sampled points by normalizing contribution magnitudes as sk=∥vk∥2/∑j∥vj∥2s_k=\|v_k\|_2/\sum_j\|v_j\|_2 and computing

    Neff=1∑ksk2.N_{\mathrm{eff}}=\frac{1}{\sum_ks_k^2}.

    This equals eight for equal contributions and approaches one when a single point dominates. We average over images within each run. Mean effective counts of 7.941–7.964 out of eight indicate broadly distributed sampled magnitudes. The directional agreement is therefore not driven by one dominant sampled point. This statistic does not characterize concentration in unsampled regions.

  • Do large contributions cancel? We compute coherence as

    C(x)=∥∑kvk∥2∑k∥vk∥2.C(x)=\frac{\|\sum_kv_k\|_2}{\sum_k\|v_k\|_2}.

    One indicates complete agreement, while values near zero indicate strong cancellation. Mean coherence is 0.927–0.940, so the aggregate retains 92.7–94.0% of the sum of individual magnitudes. The sampled contributions mostly reinforce one another despite their broadly distributed magnitudes.

  • Do different perturbation types give different directions? We compare three selection rules in the two-dimensional Penguins experiment. At radius 4 mm, the table gives median cosine between each selected point's negative margin gradient and the numerically integrated NPR gradient, pooled over six model configurations.

    Perturbation selectionCosine with the aggregate NPR gradient
    Nearest boundary point0.967
    Random boundary point0.915
    Random interior point0.588

    Both boundary selections align more closely with the aggregate than random interior points. Together with the high-dimensional results, this shows the importance of geometry and point selection. These tests compare selection rules, not noise distributions. Changing the noise density qq can also reweight the boundary integral and alter its direction.

These results identify a setting where similar directions and distributed magnitudes allow individual points to represent sampled aggregates without dominating them. They also motivate supplementation when the selected anchor is less representative. Separately, three-seed CIFAR-10/ResNet-18 controls show that adding NPR at a fixed reduced anchor coefficient improves six-shape mean PR at 4×4\times by 7.51 points (95% CI [5.35, 9.68]). We will include these analyses and training controls in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses

We thank the Area Chair and all three reviewers for their careful reading and constructive feedback. Their comments helped us clarify the contribution and strengthen the experimental evidence.

Our work asks how indicator-defined probabilistic robustness (PR) changes under model updates. Existing methods optimize surrogate losses or adapt adversarial training. These objectives do not directly characterize how PR changes under parameter updates. Under the stated assumptions, the coarea formula expresses this variation as a weighted boundary integral. This motivates studying when selected perturbations represent the aggregate direction and when supplementation is needed. A Gaussian-kernel estimator approximates this quantity in high dimensions and motivates NPR-Aug., which supplements MMA's single-anchor update.

Reviewer Z39P

  • The reviewer asked how alignment, concentration, cancellation and selection affect representativeness.

    Our results identify two relevant cases. When boundary contributions point in similar directions, an individual point can represent their aggregate without dominating it. Across the repeated CIFAR experiments, sampled-point cosines are 0.904–0.920, contribution magnitudes are broadly distributed, and coherence is 0.927–0.940, indicating little cancellation. However, a selected boundary anchor need not share this alignment. Its cosine with the same sampled aggregate is 0.545–0.748, showing that reaching the boundary alone does not ensure representativeness. The Penguins experiment illustrates a favorable setting for individual selection, with median cosines of 0.967 for the nearest boundary point and 0.915 for a random boundary point, compared with 0.588 for a random interior point. These findings motivate supplementing less representative anchors. The high-dimensional reference is a sampled aggregate, not the exact boundary integral.

Reviewer NOPR

  • The insight beyond the mathematical formulation. What is the insight beyond applying classical calculus to PR?

    Corollary 4.6 characterizes how indicator-defined PR changes under model updates. It identifies the weighted boundary contributions governing this change and motivates studying representative perturbations. We then use a smooth estimator to approximate this quantity and supplement MMA's anchor update. We acknowledge the classical foundations and have expanded the related work.

  • Simple baselines and ablations. Do gains come from the sampled objective, stochastic regularization or anchor reweighting?

    On CIFAR-10/ResNet-18, noise-CE also effectively supplements MMA. However, at 16× augmentation, NPR has higher mean PR at 4×, 8× and 16× evaluation and improves CIFAR-10-C-Full mean accuracy over matched noise-CE by 1.42 points (95% CI [0.42, 2.42]), supporting better transfer to the tested corruptions. The benefit also goes beyond anchor reweighting. At a fixed reduced anchor weight, adding NPR with 4× augmentation improves mean PR at 4× by 7.51 points (95% CI [5.35, 9.68]).

  • Geometry and aggregation. How do geometry and aggregation affect representativeness?

    Increasing the reference size does not eliminate the anchor's alignment gap. Across three MMA seeds, increasing boundary samples from 2 to 16 raises sampled-point cosine from 0.876 to 0.924, while anchor cosine remains lower at 0.688–0.711. MMA restarts give similar directions (cosines 0.945–0.984). Selection matters alongside aggregation.

  • Matched comparisons, trade-offs and reliability. Do advantages persist with matched settings and repeated runs?

    With EVaR's sampling radius matched to 4×, NPR-Aug. retains single-run PR advantages of 21.0 and 18.0 points at 8× and 16×. Increasing NPR augmentation from 4× to 16× improves PGD-10 accuracy with a modest PR decrease. The NPR/noise-CE and reduced-anchor comparisons use at least three independent training seeds.

Reviewer HDZ6

  • The reviewer requested broader evaluation and sampling, weight and aggregation ablations.

    We extended evaluation to CIFAR-100 and VGG16-bn. The three-seed CIFAR-100 results in our responses show gains over MMA of 2.38 points in clean accuracy and 1.46 in PR at 4×. Differences at larger radii are small and unresolved. On VGG16-bn, single-run PR gains over MMA are 6.84 and 5.37 points at 4× and 8×.

    Larger sampling budgets or loss weights do not necessarily improve PR. In the single-run weight scan, λ = 0.5 gives the highest observed PR at all three radii. Increasing N from four to eight adds about 15% to recorded epoch time without improving PR. Boundary diagnostics show lower anchor alignment across reference sizes, supporting the importance of point selection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.