acceptodds
Under review as a conference paper at ICLR 2027

CROCO: Certified Robustness to Optical COrruptions

Abstract

Optical systems degrade in the field through heat, mechanical decentering, and focus actuation, and recent work shows that the resulting aberrations can change the predictions of state-of-the-art image classification models. Empirical defenses reduce the risk but cannot provide guarantees. We present CROCO, the first certified robustness guarantee against physically parametrized lens aberrations. We describe each aberration by its Zernike coefficients, add Gaussian noise to these coefficients, and apply randomized smoothing. This yields an certificate in coefficient space that makes no assumption about how successive aberrations combine, so the modeled combination may be arbitrarily nonlinear. The certified radius converts into a guaranteed RMS wavefront tolerance and a range of lens sharpness (MTF), which are the quantities used to specify real lenses. The same method certifies classification, semantic segmentation, and the stability of explanations. On Imagenette, ImageNet100, and ImageNet1k, certification costs only a few points of accuracy and is nearly independent of the smoothing level. Certifying each Zernike mode separately shows that models are least robust to rotationally symmetric aberrations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Certified Robustness, Randomized Smoothing, Optical Aberrations, Adversarial Robustness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fjUxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This manuscript studies the robustness of image classification with respect to optical aberrations in camera lenses. The authors propose a mathematical framework with eight Zernike coefficients, whose randomized smoothing yields a robustness certificate. Extensive experiments, including ImageNet and several Transformers, were performed.

Critical strengths

I like the use of the optical approach with the mathematical framework, which is also appropriately connected to obtain the robustness certificate.

Critical weaknesses
  • I would like to acknowledge extensive experiments including ImageNet, but the target models are still Transformers in tiny size. More large-scale models would make this study more convincing.
  • The ALC attack is 24-dimensional, which is projected into 8-dimension in CROCO. The low flip rate in Table 18 could indicate the effect of projection itself. A more controlled ablation study is required.
  • For comparison of modes such as defocus and spherical, the same Zernike coefficients were used, but the RMS wavefront error itself could be affected by the mode choice. In other words, the comparison could also be affected by the aberration severity. This also requires more controlled experiments.
  • Eq. 3 uses Gaussian smoothing with \sigma^2 I_d, which is replaced with \sigma^2 \Pi_S in Appendix B to yield Eq. 6. The two are different things, right?
  • After Eq. 10, the authors used R=0.79 as average certified radius. However, the RMS threshold with the Marechal criterion requires a worse-case, lower bound, not the average.
  • Is the dimension-free correct? For example, squared-norm is dependent on the dimension d; a 3x dimension (from 8 to 24) would yield \sqrt{3}x increase.
  • Writing should be improved:
    • “That ellipsoid” → “That ball”
    • “Thus, amplitude based imaging would therefore be resolvable.”
    • “the whole sphere” → “the whole ball”
    • “across noise level for ResNet50” → “across noise levels for ResNet50”
    • “non adversarially” → “non-adversarially”
    • “safety critical systems” → “safety-critical systems”
    • “low dimensional set” → “low-dimensional set”
    • “pixel space certification” → “pixel-space certification” → I doubt that the authors intentionally removed all hyphens, possibly to humanize AI writing.
Questions for authors

See the weakness above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer fjUx

Thank you for your thorough review and the valuable comments on the theoretical details.

The ALC attack is 24-dimensional, which is projected into 8-dimension in CROCO. The low flip rate in Table 18 could indicate the effect of projection itself. (W2)

We agree that the projection has to be separated from the smoothing. Appendix L and the caption of Table 23 now state that this table uses the channel-averaged kernel and restrict its conclusions to this variant. Our ablation with original and averaged kernels confirms that averaging weakens the kernel. We therefore tested the attack without projection, running the original 24-dimensional ALC against a ResNet50 certified over all 24 dimensions (Imagenette, σ=1). Inside the certified radius the smoothed classifier never flips, on all 501 images, including 34 on which ALC breaks the base classifier. Outside the radius, where the certificate makes no claim, it flips the smoothed classifier on 31.8 percent of the 85 images on which ALC succeeds. Appendix L reports both.

For comparison of modes such as defocus and spherical, the same Zernike coefficients were used, but the RMS wavefront error itself could be affected by the mode choice. (W3)

Thank you for this point. A unit coefficient contributes 1/N_i waves RMS, which ranges from 1/3 for defocus to 1/8 for coma and trefoil, so equal coefficient norm does not mean equal wavefront error. Following your suggestion, we ran the controlled experiment. We certified the ResNet18 on Imagenette once per mode with noise σ_i = 0.125·N_i, which gives every mode the same wavefront error of 0.125 waves RMS. All eight modes are then certified to a tolerance between 0.343 and 0.350 waves RMS, so at equal wavefront error the model is equally robust to all modes. The finding that defocus and spherical aberration are weakest holds at equal coefficient norm. The abstract, introduction, Section 4.2 and Appendix B now state this, and Table 12 reports the new result.

Eq. 3 uses Gaussian smoothing with σ^2 I_d, which is replaced with σ^2 Π_S in Appendix B to yield Eq. 6. The two are different things, right?

Yes, they are different. N(0, σ^2 Π_S) is a degenerate Gaussian on R^d, which yields a different smoothed classifier and certifies only perturbations in the range of Π_S, which is what a per-mode certificate requires. Appendix B now makes the reduction explicit.

After Eq. 10, the authors used R=0.79 as average certified radius. However, the RMS threshold with the Marechal criterion requires a worse-case, lower bound, not the average.

You are right that the example after Equation 10 should use the radius of each image rather than the average. The guarantee itself holds per image and is not affected, only the illustration changes. The Maréchal ball is contained in an image's certificate if that image's radius is at least 8/14, and the revised Appendix F reports this per image. In the configuration with the smallest average radius, it holds for 87.4 percent of the correctly classified images and for 55.5 percent of all evaluated images.

Is the dimension-free correct? For example, squared-norm is dependent on the dimension d; a 3x dimension (from 8 to 24) would yield √3x increase.

The radius formula in Equation 5 has no explicit dependence on d, and this is what we meant by "dimension-free". However, the noise norm grows as σ√d, so p_A can change with d. In addition, equal radii are not directly comparable across dimensions. A monochromatic aberration repeats its eight coefficients in all three color channels, thus its norm is √3 times larger at d=24. A d=24 certificate of radius r therefore covers monochromatic aberrations only up to r/√3. We clarified this in Section 3, Section 4.2, Appendix C and the caption of Table 6.

I would like to acknowledge extensive experiments including ImageNet, but the target models are still Transformers in tiny size. (W1)

Certification treats the classifier as a black box, so the method places no constraint on model size, and the cost scales with the forward pass (Appendix G for the models we use). We agree that a larger model would make the study more convincing. We are therefore training and certifying a SwinV2 base on ImageNet100 with the same protocol as Table 2. After 37 of 90 epochs it reaches 73.0 percent test accuracy, and we will add its certification to the paper as soon as it finishes.

Writing should be improved [...]

Thank you for the list. All listed points are corrected, and we fixed the missing hyphens throughout.

Thank you for the detailed responses and revisions. Most of my concerns have been adequately addressed. I have only three remaining points:

  • Section 4.3 and Table 22 still seem to suggest that the 8D certificate covers the original 24D attack, although these are different settings.
  • The equal-RMS results are close to the sampling limit, making it difficult to conclude that all modes are equally robust.
  • The SwinV2-base experiment is still in progress, so this concern remains open.

Thank you for the follow-up. We uploaded a revision that addresses the three remaining points. A new Table 6 shifts the later tables by one, so the former Table 22 is now Table 23.

8D certificate and the 24D ALC attack

Section 4.3 now separates two settings. The 8D reduction of the chromatic ALC kernels is a size comparison only, since a chromatic kernel lies outside the space the 8D certificate covers, and the 69 to 84 percent is no coverage claim. The test of the original attack is the full 24D ALC against a ResNet50 certified in all 24 dimensions. Inside the radius, the smoothed classifier never flips on the 501 images, including the 34 on which ALC fools the base classifier. Appendix L and the caption of Table 23 now state this and refer to Table 25.

Equal-RMS results near the sampling limit

We agree and removed the claim of equal robustness. Raising σ does not help, because the ceiling of 3.199σ scales with it. On ImageNet1k (ResNet50, 0.125 waves, Table 6) the tolerances range from 0.275 to 0.294 waves, and every other mode lies 2 to 9 standard errors below the strongest. The flip rate 1 − n_A/n, which has no ceiling, is 0.39 percent for astigmatism and 1.49 percent for spherical aberration. Section 4.2 and Appendix B now state only this.

Larger transformer model

The SwinV2 base on ImageNet100 with σ=1 reaches acc@0 75.2 percent and ACR 2.01, against 74.4 percent and 2.00 for SwinV2 tiny, at 43.6 s per image. A newly trained ViT-B/16 reaches 77.5 percent and an ACR of 2.10. Both are in Table 2, with costs and recipes in Appendix G. The SwinV2 base runs at σ=2, σ=4 and without noise are still training and will be added to Table 2 in the final version.

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

This paper introduces CROCO, a randomized-smoothing certificate for lens aberrations parameterized by Zernike coefficients. The certificate operates in a physically meaningful parameter space and can be translated into RMS wavefront error and MTF. The authors evaluate classification, segmentation, explanation stability, per-aberration robustness, and adaptive attacks.

Critical strengths
  1. The physical parameter-space formulation is meaningful and substantially more interpretable than pixel-space robustness certificates.
  2. Connecting the certified radius to RMS wavefront error and MTF is useful for optical-system evaluation.
  3. The experiments are broad, covering multiple datasets and architectures, adaptive attacks, segmentation, and explanation stability.
  4. Per-mode certification provides useful insight into which optical aberrations are most problematic.
Critical weaknesses
  1. The theoretical construction is largely a direct application of standard randomized smoothing after treating the optical renderer and classifier as a single function. The paper also acknowledges that the same reduction underlies DeformRS. The novelty relative to existing parameter-space and transformation certificates therefore needs to be stated more precisely.

  2. The current baselines do not fully isolate the methodological contribution. In particular, comparison with a one-dimensional Gaussian-blur certificate is not sufficient because it covers a substantially weaker threat model.

  3. The optical model is simplified: it covers a limited Zernike subspace with a spatially invariant PSF and excludes higher-order defects, field-dependent blur, sensor noise, and motion blur. The resulting certificate should therefore not be interpreted as a general camera-safety guarantee.

  4. Many certified examples approach the finite-sampling ceiling, which compresses differences between models and complicates the interpretation of ACR.

  5. Some experiments require stronger controls, including the zero-aberration accuracy of each smoothing-trained model, multiple seeds, and adaptive-attack results on more than one main model/dataset setting.

Questions for authors
  1. Would applying DeformRS or another generic parameter-space smoothing method to the same Zernike renderer produce the same certificate as CROCO? If not, what is mathematically different?

  2. Can the authors compare the closest transformation-certification methods under the same renderer, architecture, training procedure, sampling budget, and confidence level?

  3. Can certified radii be compared using a common physical scale such as RMS wavefront error or MTF?

  4. How sensitive are the results to higher-order aberrations, spatially varying PSFs, and other mismatch between the renderer and a real optical system?

  5. What fraction of examples reaches the sampling ceiling in each main experiment?

2 comments
AuthorsOpenReview ↗

Response to Reviewer avxP (1/2)

Thank you for the review and the precise questions.

The theoretical construction is largely a direct application of standard randomized smoothing [...] the novelty [...] needs to be stated more precisely. (W1) Would applying DeformRS [...] produce the same certificate as CROCO? (Q1)

DeformRS (Alfarra et al., 2021) certifies deformations of pixel coordinates, I(p + v(p)), so it cannot express a lens aberration, which convolves the image with a PSF and mixes intensities instead of moving pixels. The smoothing in parameter space carries over, and so does the form of the radius. We do not claim a new bound, just as DeformRS reuses the form of the additive one. Our contribution is the perturbation class and what the guarantee offers on top of it:

  • To our knowledge, CROCO is the first exact certificate for anisotropic, physically parametrized wave aberrations.
  • The certificate has to be stated in coefficient space, because applying two aberration kernels in sequence is not the same as rendering the combined aberration (Appendix A, Figure 3).
  • Equations 9 and 10 convert the radius in closed form to a guaranteed RMS wavefront tolerance.
  • Individual Zernike modes can be certified separately (Appendix B), which has no pixel-space analogue.

Section 3 and the Related Work now state this in more detail.

The current baselines do not fully isolate the methodological contribution. [...] one-dimensional Gaussian-blur certificate is not sufficient. (W2) Can the authors compare the closest transformation-certification methods under the same renderer [...]? (Q2)

To our knowledge, no prior transformation certificate covers lens aberrations, so there is no existing method to run on our renderer. A pixel-space ℓ<sub>2</sub> certificate would have to contain the change an aberration causes across the whole image, but pixel-space smoothing certifies radii of only about 1 on ImageNet (Cohen et al., 2019), since its noise norm σ√d grows with the number of pixels (Appendix I). Blur certificates such as TSS (Li et al., 2021) cover isotropic Gaussian blur, which approximates defocus but does not contain astigmatism, coma or trefoil. We therefore use Gaussian blur as a point of reference and compare it with CROCO on a common sharpness scale in Table 16.

If you have another specific method in mind, we are happy to consider it.

Can certified radii be compared using a common physical scale such as RMS wavefront error or MTF? (Q3)

Within CROCO, the guaranteed tolerance of an image is R/8 waves RMS (Equation 10). The new Table 16 in Appendix I places the Gaussian-blur certificate, the defocus-only certificate and the full eight-mode certificate on a common scale (ResNet18, Imagenette, σ=1). For the MTF50 comparison, the eight-mode certificate is evaluated on the defocus family, since its ball contains pure defocus up to its radius. Wherever two certificates reach the same sharpness level MTF50 ≥ t, their certified accuracy differs by at most one point. Covering all eight modes costs less than one percent of the defocus tolerance and adds, on average over images, a tolerance of 0.319 waves RMS in every direction.

The optical model is simplified [...] should therefore not be interpreted as a general camera-safety guarantee. (W3) How sensitive are the results to higher-order aberrations, spatially varying PSFs, and other mismatch? (Q4)

We agree with the scope, which the Discussion and the Ethics Statement already outline. The Discussion now also states explicitly that the certificate is not a general camera-safety guarantee. It covers eight low-order modes with a spatially invariant PSF and excludes higher-order residuals, field-dependent blur, sensor noise and motion blur. We use Fringe modes 4–11, the primary aberrations of lens design. Piston and tilt are excluded because they do not blur the image. We have not measured sensitivity to higher-order modes, spatially varying PSFs, or more general renderer mismatch, and Section 5 lists the first two as limitations.

Since the proof only requires measurability in w (Appendix E) and the radius bound does not depend on d (Appendix C), higher-order modes can be added to the renderer without changing the certificate, although the RMS tolerance becomes more conservative because their normalization weights are larger. Spatially varying PSFs also fit the proof, since only the renderer changes. With P patches of independent coefficients the radius bounds the joint norm, so covering each patch up to r needs R ≥ r√P. Parametrizing the field dependence by a few low-order Zernike or Seidel coefficients avoids this and keeps d small.

AuthorsOpenReview ↗

Response to Reviewer avxP (2/2)

Many certified examples approach the finite-sampling ceiling [...] (W4) What fraction of examples reaches the sampling ceiling in each main experiment? (Q5)

Section 5 discusses the ceiling and Table 8 raises the budget to n=10^6. In addition, we measured the share per experiment from the per-image records and added it as Table 9. A correctly classified image is on the ceiling when all n samples agree. At σ=1 this share is 45.5 to 77.1 percent on Imagenette (ResNet18 71.3, ResNet50 64.6, EfficientNet B0 45.5, SwinV2 tiny 56.1, ConvNeXt tiny 77.1), 50.9 to 58.9 percent on ImageNet100, and 21.5, 23.3 and 51.2 percent for ResNet18, ResNet50 and SwinV2 tiny on ImageNet1k. Table 9 also lists σ=2 and σ=4. On Imagenette the ceiling does compress the differences between models, as you note, which is why Table 2 also reports r≥1, r≥2 and r≥3.

Some experiments require stronger controls, including the zero-aberration accuracy of each smoothing-trained model, multiple seeds, and adaptive-attack results on more than one main model/dataset setting. (W5)

Following your suggestion, we added all three controls.

  • Zero-aberration accuracy: Table 2 now reports, for every model, the accuracy of the base classifier on the aberration-free image. The certified accuracy acc@0 lies within 2.4 points of it in every configuration, so the smoothed classifier keeps the accuracy of its base classifier.
  • Multiple seeds: We trained ResNet18 and ResNet50 on Imagenette at σ=1 with three seeds each (Appendix J, Table 19). The standard deviation of acc@0 is 1.4 and 3.2 points and that of the average certified radius 0.04 and 0.12.
  • Adaptive attack on further settings: We ran the attack of Section 4.3 against ResNet18 and ConvNeXt tiny on Imagenette and ResNet50 on ImageNet100, at σ = 0.5, 1 and 2. Across 1800 further attacks, the smoothed classifier never flips inside the certified radius (Appendix K, Table 21).
Reviewer_7w4MOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CROCO applies randomized smoothing in Zernike coefficient space to certify prediction stability under modeled lens aberrations. By treating the optical renderer and classifier as one function, the construction avoids requiring successive image transformations to compose additively. The coefficient-space radius is converted into an RMS wavefront tolerance, with additional sampled MTF50 interpretations. Experiments cover classification on three datasets and several architectures, with extensions to semantic segmentation and GradCAM stability. The guarantee concerns the smoothed imaging–prediction system within the specified optical model, rather than every individual prediction of the base classifier.

Critical strengths
  1. A sound use of parameter-space smoothing for an optical threat model. Folding the renderer into the classifier gives a clear reduction to established smoothing theory without imposing an invalid image-composition assumption. The explicit Zernike normalization and RMS conversion give the radius an interpretable physical scale. This provides a useful connection between certified robustness and modeled optical tolerances.
  2. Broad experiments and informative controls. The study includes multiple architectures and noise levels, mode-removal controls, an eight- versus twenty-four-dimensional comparison, training/certification noise mismatch, and increased certification budgets. The defense-adapted coefficient-space attack is also a useful implementation check. Segmentation and explanation experiments demonstrate the construction’s breadth, while the sampling-ceiling analysis helps interpret the reported radii.
Critical weaknesses
  1. MTF50 alone does not guarantee robustness. A lens with an MTF50 value in the reported range may still fall outside the certified coefficient range. The MTF50 limits are also estimated by random sampling. Please distinguish this approximate description from the formal robustness guarantee.
  2. The ALC test changes the original attack. Table 18 averages the RGB coefficients and applies the same coefficients to all channels, removing channel differences. Please limit the conclusions to this modified attack, or evaluate the original RGB attack using the 24-dimensional setup.
Questions for authors
  1. The choice of distribution for the random aberration coefficients and its physical basis need further clarification. The authors use independent Gaussian coefficients with equal variance. Please explain whether this choice mainly supports the certification framework or also reflects differences in the strengths and correlations of aberration modes in real lenses, and clarify its relationship to real-world aberration distributions.

  2. The scope of the ALC comparison needs clarification. Appendix L states that Table 18 evaluates coefficients obtained by averaging the original coefficients across the three color channels. This modifies the original attack and may remove some channel differences. Please distinguish the original ALC attack from its channel-averaged version and revise the corresponding text and table descriptions accordingly.

  3. The relationship between MTF50 and prediction stability needs clarification. Different aberration combinations may have similar MTF50 values without all falling within the certified coefficient region. Please revise any wording implying that prediction stability is guaranteed whenever a lens’s MTF50 falls within the reported interval.

  4. Further implementation details are needed for the PSF simulation. Please specify the pupil-grid sampling density, PSF spatial extent, and image-boundary handling. Please also explain whether these settings adequately avoid PSF truncation and undersampling under strong aberrations, so that readers can assess the reliability of the simulation.

  5. Experimental hardware and computational costs need to be reported. Although the paper provides training epochs and certification sample sizes, it does not clearly specify the hardware or actual running times. Please report at least the hardware configuration and per-image certification time to help assess the method’s computational cost.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7w4M (1/2)

Thank you for the detailed and constructive review.

MTF50 alone does not guarantee robustness. A lens with an MTF50 value in the reported range may still fall outside the certified coefficient range. The MTF50 limits are also estimated by random sampling. Please distinguish this approximate description from the formal robustness guarantee. (W1, Q3)

We agree. The paper now separates the two explicitly. The guarantee covers the certified ball in Zernike coefficient space, whereas the MTF50 band is an empirical description obtained by sampling wavefronts at a given radius. The abstract, Section 4, the caption of Table 2, the Discussion, the Conclusion and Appendix F state this more explicitly. For example, an image certified at R = 1.0 keeps its prediction for every lens whose RMS wavefront error is at most 1.0/8 = 0.125 waves, that is 66.5 nm at 532 nm, whatever the mix of the eight modes is. A lens specified to the Maréchal limit of 0.071 waves is therefore covered as soon as R ≥ 8/14 ≈ 0.57, which holds for 87.4 percent of the correctly classified images in the configuration with the smallest average radius (Appendix F). MTF50 only describes what such lenses look like and is not a condition of the guarantee.

The ALC test changes the original attack. [...] limit the conclusions to this modified attack, or evaluate the original RGB attack using the 24-dimensional setup. (W2, Q2)

Thank you for pointing this out. Appendix L described the channel averaging, but the main text and the table caption did not. The caption of Table 18 (now Table 23) said that the smoothed classifier is “evaluated directly at the optimized ALC attack kernel itself”. It now says that the table evaluates the channel-averaged kernel, applied monochromatically, and Appendix L limits its conclusions to this variant. In Section 4.3 we removed the sentence “We additionally evaluate the smoothed classifier directly at the coefficient vector ALC found for each image” and the conclusion we drew from it. Section 4.3 now states the reduction from 24 to 8 dimensions and reports the 24-dimensional test instead. Following your suggestion, we evaluated the original RGB attack in the 24-dimensional setup, without any channel averaging. We ran ALC with its original 24 dimensions directly against the base classifier of a ResNet50 certified over all 24 dimensions (Imagenette, σ=1). Inside the certified radius the smoothed classifier never flips, on all 501 images, including 34 on which ALC breaks the base classifier. Outside the radius, where the certificate makes no claim, it flips the smoothed classifier on 31.8 percent of the 85 images on which ALC succeeds. An ablation on the images of Table 23 shows that the averaged kernel is weaker than the original one, so the 24-dimensional experiment is the relevant test of the original attack. Appendix L reports both.

The choice of distribution for the random aberration coefficients and its physical basis need further clarification. (Q1)

The Gaussian is the smoothing distribution of the certificate and is not meant as a model of lens statistics. The guarantee holds for every aberration inside the ball, irrespective of how lens states are distributed in practice. We added this clarification to Section 3.

Further implementation details are needed for the PSF simulation. [...] explain whether these settings adequately avoid PSF truncation and undersampling under strong aberrations [...]. (Q4)

Thank you for this suggestion. Appendix G now describes the renderer. The pupil is sampled on a 128×128 grid, with a pupil diameter of 126 samples, at 532 nm. The PSF is binned to the 3.5 μm sensor pixel with antialiased bilinear resampling and cropped to 25×25 pixels (87.5 μm), then renormalized in ℓ1.

On truncation and undersampling, Table 10 provides the numbers. At σ=1 the kernel retains 97.2 percent of the PSF energy on average and all noise draws respect the pupil sampling limit of 0.5 waves per sample. At σ=2 these values are 88.2 and 98.1 percent, and at the very strong aberrations of σ=4 they fall to 62.1 and 59.2 percent. Even at σ=4, a 75×75 window contains 96.4 percent of the energy, and with 254 pupil samples 97.4 percent of the draws respect the limit. To check whether this matters for the certificate, we certified a ResNet18 on Imagenette at σ=1 with 51×51 and 75×75 kernels instead of 25×25. The average certified radius changes from 2.53 to 2.50. Doubling the pupil sampling to 254 samples changes the rendered PSF negligibly. Appendix G now reports this.

AuthorsOpenReview ↗

Response to Reviewer 7w4M (2/2)

Experimental hardware and computational costs need to be reported. (Q5)

Appendix G now reports the hardware and computational costs. Certification ran on a single NVIDIA Tesla V100 GPU with 16 CPU cores and 20 GB of memory. Certifying one image at n_0 + n = 10,200 takes 3.4 s for ResNet18, 4.3 s for EfficientNet B0, 9.4 to 9.6 s for ResNet50, 12.4 to 12.9 s for ConvNeXt tiny and 12.7 to 17.4 s for SwinV2 tiny (median per image).

Following up on W1 and Q3, the revision adds Appendix O, which checks a real double-Gauss lens design against the per-image certificates. At the image center it is covered for 75.5 percent of the ImageNet1k images (SwinV2 tiny, σ = 1.0), nearly the 75.9 percent covered without aberration. Near the image corners it has almost the same MTF50 as a one-wave focus error (0.111 vs. 0.121 cycles per pixel), but a norm of 2.37 instead of 0.87, and is covered for 57.4 instead of 71.1 percent. Membership is thus decided by the coefficients, while MTF50 only describes lens sharpness.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Revision Summary

We would like to thank all reviewers for their careful reading and constructive feedback. We are glad that they consider our approach "a sound use of parameter-space smoothing for an optical threat model" (7w4M), find the physical formulation "meaningful and substantially more interpretable than pixel-space robustness certificates" (avxP), and like "the use of the optical approach with the mathematical framework" (fjUx). All three reviewers also acknowledge the depth of our experiments.

All additional experiments we ran in response to the reviews are finished and included in the revised paper, except the larger SwinV2 base model on ImageNet100. Its training is on track, with 73.0 percent test accuracy after 37 of 90 epochs, and we will add its certification as soon as it finishes. All substantive changes in the paper are marked in blue. New tables are placed in the appendix section they belong to, which shifts the numbering of later tables, thus Table 18 of the submission is now Table 23. An overview of the changes follows below:

Most points ask for clarifications or additional experiments, which we added, and the new experiments support the main results of the submission. Where the reviewers pointed to imprecise statements, namely the description of the attack comparison in Section 4.3, the Maréchal example after Equation 10 and the per-mode finding, we adjusted the text.

Shared remarks and how we address them:

  1. Comparison with the ALC attack (7w4M, fjUx): Section 4.3 compares our certificate with ALC (Bäuerle et al., 2026), an attack that optimizes a worst-case lens aberration with eight Zernike coefficients per color channel, 24 in total. Since our main certificate covers eight coefficients shared by all channels, Table 23 (Table 18 of the submission) averages the attack over the three channels, as Appendix L describes. Two reviewers noted that the main text did not make this clear and that a low flip rate could then come from the averaging rather than from the smoothing. We clarified Section 4.3 and the caption of Table 23 and limit the conclusions of that table to the averaged attack. To test the original attack without any averaging, we ran it against a ResNet50 certified over all 24 coefficients (Imagenette, σ=1). Inside the certified radius the smoothed classifier never changes its prediction, on all 501 images, including 34 on which the attack fools the undefended classifier. Outside the radius, where the certificate makes no claim, the same attack flips the smoothed classifier on 31.8 percent of the 85 images on which it fools the undefended one, so the attack is strong enough to find failures where they are possible. Appendix L reports this together with an ablation that measures how much the averaging weakens the attack.
  2. MTF50 and the guarantee (7w4M): Besides the radius, the paper describes certified lenses by their MTF50, the spatial frequency at which a lens transfers half of the image contrast, a common sharpness measure in lens testing. We obtain this description by sampling random wavefronts on the boundary of the certified ball and recording the range of their MTF50 values. This range describes how sharp the certified lenses typically are, but it is not a guarantee: different aberrations can have the same MTF50, so a lens whose MTF50 falls in the range can still lie outside the certified ball. The guarantee itself is stated in Zernike coefficients and implies a guaranteed tolerance in RMS wavefront error. The abstract, Section 4, the Discussion, the Conclusion and Appendix F now state this distinction explicitly.
  3. Novelty relative to DeformRS (avxP): DeformRS certifies image deformations, which move pixels, by smoothing over their parameters. We build on the same principle of smoothing in a parameter space, and our radius has the same form as in DeformRS and in standard randomized smoothing. The contribution is the perturbation class. Lens aberrations do not move pixels but change the point spread function, two aberrations cannot be composed on the image side (they are not resolvable), and no prior certificate covers them. In addition, the radius converts in closed form into a guaranteed RMS wavefront tolerance, and individual Zernike modes can be certified separately. Section 3 and the Related Work now state this explicitly.
  4. Per-mode comparison at equal wavefront error (fjUx): Section 4.2 certifies each Zernike mode separately and finds that, at the same coefficient magnitude, defocus and spherical aberration are certified to the smallest radius. As the reviewer notes, the same coefficient magnitude does not mean the same wavefront error, since the normalization differs between modes. Following the suggestion, we certified each mode with noise calibrated to the same RMS wavefront error, and every mode is then certified to a similar tolerance of 0.343 to 0.350 waves RMS, close to the maximum of 0.400 waves that the sampling budget allows, so no mode is measurably weaker at this budget. The abstract, introduction, Section 4.2 and Appendix B now state that the ordering holds at equal coefficient norm, and Table 12 reports the new result.
  5. Renderer details and sampling ceiling (7w4M, avxP): Appendix G now reports how the point spread function is rendered, including the pupil sampling, the kernel size and the share of energy the kernel retains. Certifying a retrained ResNet18 with larger kernels leaves its average certified radius nearly unchanged (2.53 against 2.50), and a finer pupil changes the rendered point spread function negligibly. Since the certified radius is bounded by the number of noise samples, Table 9 in Appendix E also reports for each model the share of images that reach this bound.
Area_Chair_NupKOpenReview ↗

Reminder: Author–reviewer discussion ends today

Dear Reviewers,

As the author–reviewer discussion period ends today (November 18, AoE), please read the authors’ responses, any relevant manuscript updates, and the other reviewers’ comments, and engage with the authors before the deadline.

In particular, please clarify which of your concerns have been addressed and which remain unresolved, and raise any remaining questions promptly so the authors have an opportunity to respond. Even if your assessment is unchanged, a brief follow-up explaining your reasoning would be helpful.

Thank you for your time and thoughtful reviewing!

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.