acceptodds
Under review as a conference paper at ICLR 2027

Randomized Manifold Smoothing for Formal Robustness to Semantically Meaningful Perturbations

Abstract

Model predictions can be sensitive to small, unrelated variations in the input, which questions the reliability of the model. Randomized smoothing provides robustness guarantees against small, adversarial perturbations, giving a certifiably robust predictor for such perturbations. However, these perturbations are applied across all directions of the ambient space, irrespective of the underlying data manifold. Consequently, its guarantees largely cover perturbations of the input that are not seen in reality rather than semantically meaningful variations in the data. To address this, we introduce Randomized Manifold Smoothing, which adapts the smoothing distribution to local data geometry by shaping Gaussian perturbations along principal directions of a locally estimated subspace. This yields anisotropic certificates over locally supported variations while retaining the statistical certification machinery of randomized smoothing. We evaluate manifold smoothing across image classification, semantic segmentation, interpretable concept representations, and vision-language retrieval tasks, including semantic distribution shifts. Across these settings, we show that manifold smoothing consistently preserves task performance while providing stronger robustness guarantees along semantically meaningful perturbations compared to standard randomized smoothing in high-dimensional data and representation spaces.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Robustness, Trustworthiness, Randomized Smoothing

Reviews

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

This paper proposes randomized manifold smoothing, which constructs a Gaussian smoothing distribution from the local PCA of neighboring training examples. The method uses whitening to obtain certificates with the covariance fixed at a reference input and analyzes how local covariance estimation affects the certification geometry. Experiments cover image classification, semantic segmentation, concept-space classification, and image-text retrieval. The authors report improved preservation of predictive performance compared with isotropic smoothing at the evaluated noise settings, together with ablations on neighborhood size, retained dimension, and sampling budgets.

Critical strengths
  1. A simple and practically useful approach to geometry-adapted smoothing.

The paper addresses a meaningful limitation of isotropic smoothing: perturbations in high-dimensional spaces can substantially disrupt the structures that a predictor relies on. Estimating the smoothing covariance through neighborhood PCA provides a concrete, training-free way to incorporate local data variation. For a fixed reference covariance and a specified support subspace, whitening connects the construction to established randomized-smoothing certification procedures. I view this combination of simplicity, applicability to existing predictors, and geometric motivation as the main reason to support acceptance.

  1. A broad empirical study with informative ablations and diagnostics.

The evaluation extends beyond a single classification benchmark to segmentation, sparse concept representations, and retrieval. The appendix examines neighborhood size, retained dimension, Monte Carlo sample count, and evaluation-set size, and reports variation across five certification seeds. The concept-activation and prediction-collapse analyses provide additional insight into how the smoothing distributions affect model behavior. While the robustness comparisons require careful interpretation, the results provide useful evidence that local covariance-based perturbations can preserve predictive structure across several settings.

Critical weaknesses
  1. The scope of the certification guarantee needs a more consistent formulation.

My positive assessment rests on interpreting the method as fixed-reference smoothing within the retained PCA subspace. The manuscript explicitly states the fixed-covariance convention and gives a supported parameterization, but some subsequent statements extend beyond what this interpretation establishes.

In particular, Eq. (9) expresses the certified region through a quadratic inequality without explicitly requiring

z−x∈span⁡(Ux,m).z-x \in \operatorname{span}(U_{x,m}).

For rank-deficient covariance, this admits unrestricted displacements in the orthogonal complement, where the smoothing argument provides no guarantee. The supported region should instead be expressed as

{x+Ux,mΛx,m1/2u:∥u∥2<r}\left\lbrace x + U_{x,m}\Lambda_{x,m}^{1/2}u : \lVert u\rVert_2 < r \right\rbrace

Theorem A.2 needs a corresponding revision: adding the support restriction alone does not establish its set inclusions when the empirical and ideal subspaces differ. This is a substantive issue in the set-level analysis, rather than merely a typographical omission.

The fixed-reference predictor also needs to be distinguished from the query-adaptive predictor. Certifying gx(z)g_x(z) with covariance fixed at xx does not automatically certify g(z)=gz(z)g(z)=g_z(z), which recomputes the covariance at each query. The intended inference procedure and the scope of the guarantee should therefore be stated consistently.

I regard these as potentially addressable limitations of the current formulation. They do not negate the useful fixed-reference, subspace-restricted construction, but they do limit the broader formal claims.

  1. The experimental advantage should be interpreted with the different perturbation scales and geometries in mind.

Using the same scalar σ\sigma does not match perturbation magnitude between the methods:

E[∥δiso∥22]=σ2d.\mathbb{E}\left[\|\delta_{\mathrm{iso}}\|_2^2\right]=\sigma^2 d.

E[∥δman∥22]=σ2tr⁡(Σxloc).\mathbb{E}\left[\|\delta_{\mathrm{man}}\|_2^2\right]=\sigma^2\operatorname{tr}\left(\Sigma_x^{\mathrm{loc}}\right).

Consequently, improved prediction preservation can reflect both covariance orientation and a smaller effective perturbation. The authors acknowledge that radii and volumes in the different geometries are not directly comparable, and Appendix D provides additional radius-threshold curves. Nevertheless, the reported comparisons do not fully isolate the contribution of local geometry or establish superiority under a shared semantic perturbation requirement.

Similarly, certification at examples from an attribute-defined distribution shift does not establish that the attribute change itself is covered by the certificate. The retrieval results also provide empirical robustness evidence rather than certificates against the caption corruptions. These limitations qualify the interpretation of the results, although the observed predictive stability under the proposed noise model remains useful.

Questions for authors
  1. Please clarify the exact certified set in the rank-deficient case. Does the implementation restrict noise to the retained PCA subspace, or add positive variance in every ambient direction? How should Eq. (9) and Theorem A.2 be interpreted consistently with Eq. (10), Algorithm 1, and the eigenvalue floor in Appendix B? In particular, how are the set-level statements handled when the empirical and ideal support subspaces differ?

  2. Is the intended deployed predictor the reference-conditioned predictor gx(z)g_x(z), or the query-adaptive predictor g(z)=gz(z)g(z)=g_z(z)? For the former, how is the reference input selected or retained during inference? For the latter, what additional argument establishes robustness when the covariance changes with the query?

  3. Can the authors quantify the perturbation magnitudes used in the main comparisons, for example through covariance traces or noise-norm statistics? How much of the observed advantage remains when the isotropic baseline is compared at a matched perturbation magnitude, and what common robustness requirement is intended when interpreting the certified-accuracy differences?

  4. In the attribute-defined OOD experiments, does the nearest-neighbor reference pool contain only the classifier-training subgroup with a=0a=0, or also examples with a=1a=1? Do these experiments establish coverage of the attribute change itself, or certification of additional local perturbations around the shifted examples?

These questions concern the interpretation and validity of the submitted contribution, rather than unrelated future extensions. Clear answers would strengthen my confidence in the positive recommendation.

4 comments
AuthorsOpenReview ↗

Rebuttal Response to Reviewer DRUj (1/2)

Dear Reviewer DRUj,

Thank you for the thorough and constructive review. We are glad you see simplicity, applicability to existing predictors, and geometric motivation as reasons to support acceptance, and that local covariance-based perturbations preserve predictive structure across settings. We address your questions below.

1. Scope of the certified region (W1, Q1)

Certified region. We agree: Eq. 9 must restrict z−xz-x to span(Ux,m)\mathrm{span}(U_{x,m}), the retained subspace where Eq. 11 and Algorithm 1 sample noise:

Ex(r)={x+Ux,mΛx,m1/2v: v∈Rm, ∥v∥2≤r}\mathcal{E}_x(r)=\lbrace x+U_{x,m}\Lambda_{x,m}^{1/2}v:\ v\in\mathbb R^m,\ \lVert v\rVert_2\le r\rbrace

Without it, the quadratic inequality defines an unbounded cylinder when (m<d)(m<d). The guarantee covers perturbations within the retained subspace that satisfy the bound, but does not extend to discarded directions.

Theorem A.2. We agree that the support restriction alone does not establish the inclusions when the empirical and ideal subspaces differ. For Theorem A.2, the stated inclusions hold for the ambient quadratic sublevel sets, but when span⁡(U^x,m)≠span⁡(Ux⋆)\operatorname{span}(\widehat U_{x,m})\ne\operatorname{span}(U_x^\star), they do not imply literal inclusions between the support-restricted certified ellipsoids. These ellipsoids can instead be compared after aligning their subspaces, with geometric convergence following from covariance convergence. We will add the restriction to Eq. 9 and restate Theorem A.2 for the support-restricted sets.

Eigenvalue floor (App. B). Flooring only replaces each retained eigenvalue by max⁡{λ^x,j,10−6}\max\lbrace\widehat\lambda_{x,j},10^{-6}\rbrace for numerical stability, consistently in sampling and certification. Theorem A.2 applies unchanged when the floor is inactive; otherwise it must account for the flooring error or compare against the floored ideal covariance.

Implementation. Noise is restricted to the retained PCA subspace; no variance is added in other ambient directions. Our main experiments retain m=min⁡(k,d)=500m=\min(k,d)=500 components from k=500k=500 neighbors; we will state mm explicitly in Sec. 4.

2. Matched perturbation magnitude (W2, Q3)

We agree that the same σ\sigma does not imply the same magnitude: E∥δISO∥22=σ2d\mathbb E\lVert\delta_{\mathrm{ISO}}\rVert_2^2=\sigma^2d, whereas E∥δRMS∥22=σ2Tx\mathbb E\lVert\delta_{\mathrm{RMS}}\rVert_2^2=\sigma^2T_x with Tx=tr⁡(Λ~x,m)T_x=\operatorname{tr}(\widetilde\Lambda_{x,m}). To isolate geometry from magnitude, we add two Gaussian baselines matching Randomized Manifold Smoothing’s (RMS) expected squared norm:

  • Matched ambient ISO (new) : Gaussian noise across all d ambient dimensions, with σI2=σ2tr⁡(Λ~x,m)/d\sigma_{\mathrm I}^2=\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m})/d, matching RMS's expected squared perturbation norm.

  • Matched projected ISO (new): Gaussian noise restricted to the same retained PCA subspace as RMS, with α2=σ2tr⁡(Λ~x,m)/m\alpha^2=\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m})/m, matching the expected squared norm while assigning equal variance to all retained directions.

Obsereved Perturbation magnitude (E∥δ∥22\sqrt{\mathbb{E}\|\delta\|_2^2}):

Methodσ=.1σ=.5σ=.7σ=1σ=2
ISO38.80193.99271.58387.98775.96
RMS6.05±.7530.25±3.7742.35±5.2860.50±7.53121±15.06

Certified accuracy (%):

Methodσ=.1σ=.5σ=.7σ=1σ=2
RMS (ours)9291908672
Matched projected ISO9289878269
Matched ambient ISO9280684745
Standard ISO7245454545

Results. Matching magnitude explains part of the gap to standard ISO, but RMS substantially outperforms matched ambient ISO at larger scales. Restricting noise to the PCA subspace helps further, and RMS matches projected ISO at small scales and outperforms it at larger ones through eigenvalue-dependent scaling. RMS's advantage thus comes not only from smaller magnitude or dimensionality reduction, but also from aligning perturbation variance with local data geometry. Complete formulations and results are added in Appendix D.1.

AuthorsOpenReview ↗

Rebuttal Response to Reviewer DRUj (2/2)

3. OOD and scope of certificates (W2, Q4)

The nearest-neighbor pool contains only the classifier-training subgroup (a=0a=0), so the experiments are truly OOD: geometry around each shifted example is estimated from in-distribution data. We agree: they certify local perturbations around shifted examples, not the attribute change itself. They still answer a key question: do guarantees over semantic variations persist on data unseen in training? Even with in-distribution neighbors only, RMS keeps nontrivial certified accuracy across all six shifts (Fig. 3), whereas isotropic smoothing degrades. Fig. 4 shows that certified variations include attribute-like changes (mouth opening, smiling), though they remain local. We will rephrase Sec. 5.1 to state this scope precisely.

Summary. We thank the reviewer for their suggestions: their comments sharpen the formal scope of our guarantee (Eq. 9, Theorem A.2) while leaving the method and its results unchanged. We have incorporated the changes in our manuscript, addressing the certification region (Eq. 9 updated), the noise magnitude comparison in App. D.1, and clarified the OOD sampling pool in Sec. 5. RMS remains a simple, training-free way to certify existing predictors along data-supported directions, and the new matched-magnitude baselines show that our advantage holds beyond noise scale and comes from aligning the noise with local data geometry.

We hope these clarifications and additions will convince you to give us a positive recommendation.

Reviewer_DRUjOpenReview ↗

Response to rebuttal

Many thanks to the authors for their detailed rebuttal. I am satisfied with how they addressed two of my main concerns.

  • Scope of the certificate. The revised Eq. (9) explicitly restricts the certified region to the retained PCA subspace, and Eq. (10) clarifies that the covariance is held fixed at the reference input.
  • Matched-noise comparison. The authors added the projected isotropic baseline I was hoping to see in Table 6. It matches both the PCA subspace and the expected squared perturbation norm, making the benefit of eigenvalue-dependent scaling at larger noise levels clearer.

I am therefore happy to raise my score from 2 to 3. Best of luck to the authors!

AuthorsOpenReview ↗

Response to Reviewer DRUj

We thank reviewer DRUj greatly for staying engaged and acknowledging our detailed rebuttal. We very much appreciate raising their score from 2 to 3 (weak accept).

If there are still weak points open that the reviewer might think render the current manuscript not a clear acceptance (rating of 4), please let us know. We would highly value your further constructive input to improve the manuscript. We are happy to address any remaining such points while the discussion period is open.

Reviewer_4oZoOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Randomized smoothing (RS) is a ML technique which adapts an underlying classifier f(x)f(x) into smoothed version g(x)=Eδxf(x+δx)g(x)=E_{\delta_x}f(x+\delta_x) by taking average over gaussian perturbations around the original input δx∼N(0,σ2I)\delta_x\sim N(0,\sigma^2I). This paper extends this idea to randomised manifold smoothing (RMS), where the perturbations are not isotropic but an-isotropic gaussian perturbations δx∼N(0,σ2Σxloc)\delta_x\sim N(0,\sigma^2\Sigma^{loc}_{x}) depending on the local data structure. The perturbations are transformed accordingly to the local geometry or curvature of the data. This enables this perturbations to be semantically meaningful. The paper establises robustness certifates for RMS akin to the original RS setting, moreover empirically validate superiority of these perturbations in maintaining higher accuracy at higher σ\sigma compared to the isotropic counterpart.

Critical strengths
  1. Intuitive generalization of RS to account for data curvature: It has been well known that random perturbations alogn all data direction lose interpretability and don't account for the natural data variation. This has plafgued the original RS formulation with the question are these perturbations truly adversarial in a semantic sense. This paper addresses that qualm, the perturbations in this framework make sense semantically. This has been illustrated explicitly with experiments Figure 4,5,6.
  2. Huge improvements in accuracy while retaining task utility: Table 1, Table 3 show that across a variety of tasks segmentation, classification, semantic retrieval the robustness accuracy consistently outperforms the isotropic variant.
  3. OOD robustness guarantee: RMS is resiliant to larger σ\sigma values even if trained on smaller ones. Section 5.1.
Critical weaknesses

Overall, I think the paper's strengths outweight the weaknesses, however few points that should be clarified to strengthen the message:

  1. Pertubation magnitudes might not be comparable directly? :The authors should clarify that even though σ\sigma might be large the actual pertubation magnitude can be much smaller and potentially be contributing to a lot of resilence. That is ∣∣δiso∣∣≈σ⋅d||\delta_{\rm iso}||\approx \sigma \cdot d, here dd is the dimension of the manifold; but for RMS ∣∣δaniso∣∣≈σ⋅Tr(Σxloc)||\delta_{\rm aniso}||\approx \sigma \cdot Tr(\Sigma_{x}^{loc}) which could have much smaller magnitude. In figure 7, we actually see this the ℓ2\ell_2 magnitude for isotropic pertubations is much larger (90 vs 5.).
  2. Computing Σxloc\Sigma_x^{loc} needs enough neighbours to reflect the true curvature, most image data have much lower dimension than the ambient dimension so the current 500 NN is probably good enough. But a small discussion for the accurate estiamtion and the intrinsic dimension of the data manifold would be desirable.
Questions for authors

Refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Rebuttal Response to Reviewer 4oZo

Dear Reviewer 4oZo,

Thank you for your positive assessment of our work, particularly for recognizing the intuitive motivation of data-aligned smoothing providing semantically related perturbations rather than adversarial and acknowledging the huge improvements in the accuracy of appreciating our method while retaining task utility performance across task and domain and as well as the OOD resilience. We appreciate your suggestions that attention to the role of intrinsic dimension needs to be highlighted. We therefore collectively address the concerns below.

1. Perturbation magnitude (W1-W2): We agree that equal noise σ\sigma does not imply equal perturbation magnitudes:

E∥δISO∥22=σ2d,E∥δRMS∥22=σ2tr⁡(Λ~x,m).\mathbb{E}\|\delta_{\mathrm{ISO}}\|_2^2=\sigma^2d,\qquad \mathbb{E}\|\delta_{\mathrm{RMS}}\|_2^2=\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m}).

This difference is inherited in Randomized Manifold Smoothing (RMS) method that shaping noise according to local covariance rather than distributing it uniformly across ambient directions.

To isolate noise magnitude, subspace restriction, and eigenvalue-dependent scaling, we compare four schemes:

  1. Standard ISO (existing): Uniform Gaussian noise across all dd dimensions at the same nominal.
  2. Matched ambient ISO (new): Matches RMS's expected squared perturbation norm across all dd dimensions, controlling for magnitude.
  3. Matched projected ISO (new): Matches the norm within the same mm-dimensional PCA subspace as RMS, but distributes variance uniformly, isolating eigenvalue-dependent scaling.
  4. RMS (ours, existing): Noise in the retained PCA subspace, with variance proportional to the local eigenvalues.

The new baselines use:

σI2=σ2tr⁡(Λ~x,m)d,α2=σ2tr⁡(Λ~x,m)m.\sigma_{\mathrm I}^2=\frac{\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m})}{d},\qquad \alpha^2=\frac{\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m})}{m}.

Observed Perturbation magnitude (E∥δ∥22\sqrt{\mathbb{E}\|\delta\|_2^2}):

Methodσ=.1σ=.5σ=.7σ=1σ=2
ISO38.80193.99271.58387.98775.96
RMS (Ours)6.05±.7530.25±3.7742.35±5.2860.50±7.53121±15.06

Certified accuracy (%):

Methodσ=.1σ=.5σ=.7σ=1σ=2
RMS (ours)9291908672
Matched projected ISO9289878269
Matched ambient ISO9280684745
Standard ISO7245454545

Results. Matching perturbation magnitude explains part of the gap between standard isotropic smoothing and RMS. However, RMS substantially outperforms matched ambient isotropic smoothing at larger noise scales. Restricting perturbations to the PCA subspace provides further improvements, while RMS matches at decent perturbation; however, it outperforms projected isotropic smoothing through eigenvalue-dependent scaling at larger perturbation. Thus, RMS's advantage is not solely due to smaller noise magnitude or dimensionality reduction, but also to aligning perturbation variance with local data geometry.

Complete formulations and results are provided in Appendix D.1.

2. Intrinsic Dimension Estimation (W4):

Thank you for highlighting intrinsic dimensionality in local manifold estimation. We agree with your observation that data in domains such as computer vision and language often exhibit substantially lower effective dimensionality than their ambient feature spaces, making local covariance estimation with a moderate number of neighbors more feasible.

For data with higher intrinsic dimensionality, more neighbors may be required to reliably estimate the local covariance and principal directions. From a practical perspective, when the true intrinsic dimension is unknown, the effective rank of the local covariance can provide a proxy for assessing local dimensionality, and if the estimated dimension is high, a larger and sufficiently dense reference dataset may be necessary to obtain reliable estimates. When such a data estimate is unavailable or unreliable, standard isotropic randomized smoothing remains a reasonable fallback, as it does not require estimating local manifold geometry. We have incorporated this meaningful discussion into the manuscript. We thank the reviewer for pointing out this important consideration.

Summary. We added two matched-magnitude baselines (App. D.1) and expanded the discussion of intrinsic dimension (Sec. 6). These changes and points strengthen our manuscript by showing that RMS's advantage comes from aligning noise with local data geometry, beyond a smaller noise magnitude, and by clarifying when the method applies. We thank the reviewer for making these improvements possible.

We hope these additions address your concerns and would be happy to discuss further during the discussion period.

Reviewer_4oZoOpenReview ↗

Reply to rebuttal

I thank the authors for the extra experiments directly addressing my questions, I would like to maintain my positive score.

Reviewer_f1d9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Randomized Manifold Smoothing, a data-adaptive extension of randomized smoothing that aligns the smoothing distribution with the local geometry of the data. Instead of adding isotropic Gaussian noise in all ambient-space directions, the method estimates a local covariance using nearest neighbors, extracts the corresponding PCA directions, and scales Gaussian perturbations according to the observed local variation. Through a whitening transformation, the authors connect this anisotropic smoothing procedure to the standard randomized-smoothing framework and obtain manifold-aligned certified regions. The paper evaluates the method across a diverse set of settings: image classification, semantic segmentation, sparse interpretable concept representations, and vision-language retrieval. Across these tasks, manifold smoothing substantially preserves predictive performance and semantic structure at perturbation levels where standard isotropic smoothing deteriorates. Overall, I find the paper interesting and timely. The central idea is simple, intuitive, and broadly applicable, and the empirical evidence is strong enough to support the main qualitative conclusion that smoothing along locally supported directions can be considerably more useful than spreading perturbation mass uniformly throughout a high-dimensional ambient space.

Critical strengths
  1. The paper has a strong motivation and problem formulation. The paper clearly identifies the mismatch between isotropic randomized smoothing and low-dimensional, anisotropic natural data variation.
  2. The method is simple and intuitive. Local kNN covariance and PCA are used to shape Gaussian noise along data-supported directions, making the approach easy to understand and implement.
  3. The method is tested on classification, segmentation, interpretable concept spaces, and vision-language retrieval across multiple representation spaces.
  4. The paper presents a useful theoretical analysis. The appendix provides convergence results showing when the estimated local covariance and PCA subspace approximate the true tangent geometry.
Critical weaknesses
  1. The certificate definition needs clarification. The proof keeps the covariance fixed at the original input, while the method is described using an input-dependent covariance; the exact certified predictor should be stated more precisely.
  2. The low-rank certificate could be clearer. When only (m<d) PCA directions are retained, the certified region should explicitly be described as restricted to the retained PCA subspace.
  3. Same sigma comparison is not fully controlled. Isotropic and manifold noise can have very different effective magnitudes and dimensionalities, so matched-energy or matched-norm baselines would strengthen the conclusions.
  4. A projected-isotropic or local-PCA isotropic baseline would better isolate the benefit of anisotropic eigenvalue scaling from dimensionality reduction.
Questions for authors
  1. Can you clarify the exact classifier being certified? In particular, is the local covariance recomputed at a perturbed point (z), or is the covariance estimated at the original input kept fixed throughout certification?
  2. How should the certificate be interpreted when only (m<d) PCA directions are retained? Is the certified region explicitly restricted to the retained PCA subspace?
  3. Can you provide a matched-perturbation baseline? For example, isotropic smoothing in the same PCA subspace or comparisons with equal expected perturbation norm would better isolate the benefit of manifold alignment.
2 comments
AuthorsOpenReview ↗

Rebuttal Response to Reviewer f1d9

Dear Reviewer f1d9,

Thank you for your constructive review and for recognizing the intuitive motivation behind manifold-aligned smoothing for providing semantic perturbation guarantees, the breadth of our results across tasks (classification, segmentation, multimodal alignment, and space (pixel and latent space)), and theoretical analysis of tangent geometry estimation. We appreciate your suggestions regarding noise magnitude comparisons and certification scope, which we address below.

1. Certified Classifier and Fixed Reference Covariance (W1, Q1)

Thank you for pointing out the missing parameterization. We clarify that the certified classifier is the fixed-reference predictor:

gx(z)=arg⁡max⁡c∈YPr⁡δ∼N(0,σ2Σxloc)[f(z+δ)=c].g_x(z)=\arg\max_{c\in\mathcal{Y}}\Pr_{\delta\sim\mathcal{N}(0,\sigma^2\Sigma_x^{\mathrm{loc}})}[f(z+\delta)=c].

The local covariance is estimated once at the reference input x and held fixed throughout certification, rather than recomputed at z. This is already described in Sec. 3.3, following Eq. 9, and is consistent with Algorithm 1.

Our guarantee establishes: gx(z)=gx(x)g_x(z)=g_x(x) within the certified region Ex(rwhite)\mathcal{E}_x(r_{\mathrm{white}}). We will explicitly introduce gxg_x in Eq. 4.

2. Certified Region under Low-Rank PCA (W2, Q2)

Yes, when m<d, the certificate covers perturbations only within the retained PCA subspace span⁡(Ux,m)\operatorname{span}(U_{x,m}), not arbitrary ambient-space displacements. As specified in Eq. 11 and Algorithm 1, noise is sampled along the retained m principal directions.

We will revise Eq. 9 to explicitly state:

Ex(r)={x+Ux,mΛx,m1/2v:∥v∥2≤r}\mathcal{E}_x(r)=\{x+U_{x,m}\Lambda_{x,m}^{1/2}v:\|v\|_2\leq r\}

3. Matched Perturbation Magnitude and Projected-Isotropic Baselines (W3–W4, Q3)

We agree that equal nominal does not imply equal perturbation magnitudes:

E∥δISO∥22=σ2d,E∥δRMS∥22=σ2tr⁡(Λ~x,m).\mathbb{E}\|\delta_{\mathrm{ISO}}\|_2^2=\sigma^2d,\qquad \mathbb{E}\|\delta_{\mathrm{RMS}}\|_2^2=\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m}).

This difference follows from Randomized Manifold Smoothing (RMS) (ours) shaping noise according to local covariance rather than distributing it uniformly across ambient directions.

To isolate noise magnitude, subspace restriction, and eigenvalue-dependent scaling, we compare four schemes:

  1. Standard ISO (existing): Uniform Gaussian noise across all d dimensions at the same nominal.
  2. Matched ambient ISO (new): Matches RMS's expected squared perturbation norm across all d dimensions, controlling for magnitude.
  3. Matched projected ISO (new): Matches the norm within the same m-dimensional PCA subspace as RMS, but distributes variance uniformly, isolating eigenvalue-dependent scaling.
  4. RMS (ours, existing): Noise in the retained PCA subspace, with variance proportional to the local eigenvalues.

The new baselines use:

σI2=σ2tr⁡(Λ~x,m)d,α2=σ2tr⁡(Λ~x,m)m.\sigma_{\mathrm I}^2=\frac{\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m})}{d},\qquad \alpha^2=\frac{\sigma^2\operatorname{tr}(\widetilde{\Lambda}_{x,m})}{m}.

Observed Perturbation magnitude (E∥δ∥22\sqrt{\mathbb{E}\|\delta\|_2^2}):

Methodσ=.1σ=.5σ=.7σ=1σ=2
ISO38.80193.99271.58387.98775.96
RMS6.05±.7530.25±3.7742.35±5.2860.50±7.53121±15.06

Certified accuracy (%):

Methodσ=.1σ=.5σ=.7σ=1σ=2
RMS (ours)9291908672
Matched projected ISO9289878269
Matched ambient ISO9280684745
Standard ISO7245454545

Results. Matching perturbation magnitude explains part of the gap between standard isotropic smoothing and RMS. However, RMS substantially outperforms matched ambient isotropic smoothing at larger noise scales. Restricting perturbations to the PCA subspace provides further improvements, while RMS matches at decent perturbation, but outperforms projected isotropic smoothing through eigenvalue-dependent scaling at larger perturbation. Thus, RMS's advantage is not solely due to smaller noise magnitude or dimensionality reduction, but also to aligning perturbation variance with local data geometry. Complete formulations and results are provided in Appendix D.1.

Summary. We have clarified that the certified classifier uses covariance estimated once at the reference input and held fixed (Eq. 4 updated), and that the certificate is restricted to the retained PCA subspace (Eq. 9 updated). We have updated the manuscript accordingly. We also added two matched-magnitude isotropic baselines to address noise magnitude differences (App. D.1 and discussion added). The results show that RMS benefits from both subspace restriction and eigenvalue-dependent anisotropic scaling, beyond noise magnitude alone.

We hope these additions address your concerns and would be happy to discuss further during the discussion period.

AuthorsOpenReview ↗

Rebuttal Response Follow-up to Reviewer f1d9

Dear Reviewer f1d9,

We hope that our rebuttal and revised manuscript have addressed your concerns.

Given the short remaining time, we would be grateful if you could give us feedback if there is any concerns remain open. We would be happy to provide further clarification while the discussion period is open.

Thank you again for your valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.