acceptodds
Under review as a conference paper at ICLR 2027

ContRS: Generalizing Randomized Smoothing to Continuous Outputs

Abstract

Randomized smoothing certifies adversarial robustness by aggregating a base model’s predictions under injected noise. In classification, its classical formulation compares class probabilities under Gaussian perturbations to certify exact label invariance. For continuous outputs, however, model predictions can remain close without being identical, making a quantitative bound on prediction change more appropriate. Existing extensions typically rely on task-specific constructions and do not directly map a prescribed output tolerance to a certified input radius across different output types. To this end, we introduce ContRS, a new unified framework that defines the smoothed prediction as the center of an output neighborhood with maximum probability mass and recovers classical randomized smoothing as a special case. At the population level, we show that a probability margin over competing neighborhoods yields a certified radius within which the smoothed prediction changes by at most the prescribed tolerance. For metric and symmetric relaxed-triangle discrepancies, we derive finite-sample certificates using a center selected independently of the certification samples. For general discrepancies, we provide empirical robustness estimates based on sampled competitors. Across diverse continuous-output tasks, ContRS yields larger certified input radii through output-region selection at matched output-region scales and tighter output bounds at matched input budgets, with lower or comparable smoothing error compared with representative baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Certified Robustness, Randomized Smoothing, Continuous Outputs

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_QyJYOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper claims a way (ContRS) to extend randomized smoothing from categorical model outputs to continuous-output tasks. It selects the center of an output neighborhood with the largest probability mass and derives a certified input radius for a prescribed output tolerance. Experiments compare ContRS with center smoothing and safe region certification across several prediction tasks.

Critical strengths
  • ContRS considers several output distances, making the method applicable across structured-output tasks.
Critical weaknesses
  • The claimed research gap in the introduction part is overstated. For example, existing percentile-smoothing bounds can be inverted to obtain a certified input radius for a prescribed output tolerance. Given certified bounds Lx(R)≤g(x+δ)≤Ux(R)L_x(R)\le g(x+\delta)\le U_x(R) for ∣∣δ∣∣2<R||\delta||_2<R, any RR satisfying Lx(R)≥g(x)−rL_x(R)\ge g(x)-r and Ux(R)≤g(x)+rU_x(R)\le g(x)+r certifies the desired tolerance rr. Likewise, the center-smoothing guarantee gives R=σΦ−1(pr)R=\sigma\Phi^{-1}(p_r), where pr=Pr⁡[f(x+ϵ)∈B(gcs(x),r/(2κ))]>1/2p_r=\Pr[f(x+\epsilon)\in B(g_{\mathrm{cs}}(x),r/(2\kappa))]>1/2, since this ball retains majority mass under perturbation and the center-smoothing bounds the output change by 2κ⋅r/(2κ)=r2\kappa\cdot r/(2\kappa)=r. According to these, the stated certification problem is a straightforward corollary of existing smoothing guarantees. Thus I find the claimed research gap unconvincing.

  • The theoretical novelty beyond existing smoothing certificates is insufficiently established. In the metric and relaxed-triangle settings, ContRS reuses the Gaussian radius formula of safe region, and the paper attributes its improvements to output-region selection. This further suggests that the contribution lies in the predictor and region-selection strategy rather than a new radius calculation.

Questions for authors
  1. Why does directly inverting existing percentile- or center-smoothing bounds not solve your stated research problem?

  2. Does ContRS still outperform safe region when the target probability of the baseline is tuned rather than fixed at p=0.55p=0.55?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Comments (Part 1)

Thank you for your time and feedback on our work. We have provided detailed, point-by-point responses to the issues and questions you raised. We have also added clarifications in Introduction and at the end of Method in the revised PDF. We hope these responses and revisions address your concerns and clarify our motivation and contributions.

Q1: The claimed research gap in the introduction part is overstated. The stated certification problem is a straightforward corollary of existing smoothing guarantees.
Re: Thank you for your suggestion. We would like to further clarify the motivation of our work. Our key point is not that existing extensions of randomized smoothing cannot derive certified input radii by inverting or converting their guarantees. Rather, their output representations, aggregation criteria, and certification targets can impose limitations on cross-task applicability and certificate tightness. These limitations are not removed simply by inverting an existing bound.

For percentile smoothing, handling variable-length or unordered structured outputs typically requires constructing a consistent coordinate representation and translating coordinate-wise intervals into a bound on overall output variation. This requires task-specific design and analysis. Moreover, relying on coordinate-wise bounds may overlook joint constraints among output components, leading to conservative variation bounds and smaller certified input radii.

For center smoothing, inversion can provide a certified input radius for a prescribed output tolerance, but it does not change the prediction center. The center is still chosen as the center of the smallest ball containing at least half of the output probability mass. When the ball radius is fixed according to the prescribed tolerance, this center may capture less probability mass than another center. Consequently, the corresponding majority-based certificate is sub-optimal.

Thus, the key distinction lies not merely in deriving a certified input radius from an output tolerance, but in incorporating the prescribed tolerance into prediction aggregation and certification analysis to obtain tighter certificates. Guided by this principle, ContRS extends randomized smoothing to continuous outputs and offers a new approach to robustness certification that goes beyond a straightforward corollary of existing smoothing guarantees.

Q2: The theoretical novelty beyond existing smoothing certificates is insufficiently established. This further suggests that the contribution lies in the predictor and region-selection strategy rather than a new radius calculation.
Re: Thank you for this comment. We would like to clarify that the Gaussian radius formula is an established tool in classical randomized smoothing for deriving certified input radii from probability bounds. Both center smoothing and safe region build on Gaussian probability bounds derived from the Neyman–Pearson lemma [1], using their respective prediction constructions and certification analyses to establish robustness guarantees. Sharing this common analytical tool therefore does not, by itself, imply a lack of theoretical novelty. Building on the same foundation, our main theoretical contribution is to incorporate the prescribed output tolerance directly into the definition of the smoothed prediction and establish the corresponding population stability conditions and finite-sample certification guarantees, as detailed below.

First, definition of the smoothed prediction. ContRS incorporates the prescribed output tolerance into its prediction definition to improve certificate tightness: it sets the neighborhood radius according to the allowed output variation, then selects the center whose neighborhood has the largest probability mass. This construction extends the single-label events used in classical classification smoothing to events defined by output neighborhoods, aggregating probability mass across nearby outputs to define predictions in continuous output spaces. For discrete labels equipped with the discrete distance, these neighborhoods reduce to single-label events, recovering classical randomized smoothing. Thus, the output tolerance directly informs prediction aggregation and provides the basis for subsequent stability analysis and certification.

AuthorsOpenReview ↗

Response to Reviewer Comments (Part 2)

Second, certification via competing neighborhoods. Theorem 4.1 treats outputs whose discrepancy from the original prediction exceeds the prescribed tolerance as competing centers. It compares the probability mass of the original prediction’s neighborhood with the supremum of the masses of all competing neighborhoods. When the former exceeds the latter, this probability gap yields a certified input radius within which the prediction recomputed after an input perturbation remains within the prescribed tolerance of the original prediction. This analysis extends class-probability comparisons in classical random smoothing to neighborhood-probability comparisons for continuous outputs. It does not require the original neighborhood to contain more than half of the probability mass, and the discrepancy need not be symmetric or satisfy a triangle inequality.

Third, finite-sample certification guarantees. Corollary 4.2 shows that a center selected from finitely many samples can support valid certification of the population smoothed prediction defined by maximizing neighborhood probability mass, even when the selected center is not globally optimal over the output space. Specifically, center selection and probability estimation use independent samples. If the lower confidence bound on the selected neighborhood’s probability mass exceeds one half, the relevant geometric conditions yield a certified input radius. With the prescribed confidence, the following holds simultaneously for every perturbation within that radius: the perturbed population prediction remains within tolerance of both the original population prediction and the selected sample center. This guarantee requires neither exact optimization over the entire output space nor convergence of the sampled center to a population optimum; it follows from the probability mass of the selected neighborhood.

[1] Cohen, J., Rosenfeld, E., Kolter, J. Z. Certified Adversarial Robustness via Randomized Smoothing. ICML, 2019.

Q3: Why does directly inverting existing percentile- or center-smoothing bounds not solve your stated research problem?
Re: Thank you for this question. Direct inversion can indeed yield a certified input radius for a prescribed output tolerance in the settings covered by existing percentile and center smoothing guarantees. The distinction we intend to emphasize concerns how the smoothed prediction is constructed and certified, rather than whether an input radius can be obtained by inversion.

For percentile smoothing, inversion directly provides the desired tolerance certificate for scalar outputs. For vector outputs, it also applies when the coordinate-wise bounds can be combined into a valid bound under the desired output discrepancy. However, this requires relating the coordinate representation and its interval bounds to that discrepancy. For variable-length or unordered structured outputs, constructing a consistent coordinate representation and deriving such a bound may require task-specific design and analysis [2]. Moreover, relying on coordinate-wise intervals may overlook joint constraints among output components, leading to conservative variation bounds and smaller certified input radii. Inverting the bounds does not remove these representation requirements or make the bounds on output variation tighter. ContRS instead selects the center whose output neighborhood has the largest probability mass, with the neighborhood defined by the chosen discrepancy and prescribed tolerance. This aligns prediction aggregation with the allowed output variation and enables certification to directly bound the discrepancy between the original and perturbed smoothed predictions.

For center smoothing, inversion can yield a valid certified input radius, but it retains the center of the smallest ball containing at least half of the output probability mass. This center need not capture the greatest probability mass at the neighborhood radius determined by the prescribed output tolerance, which can limit the resulting certified input radius. In contrast, ContRS fixes the neighborhood radius according to that tolerance and selects the center with the largest neighborhood probability mass.

Specifically, for the same base predictor, noise level, discrepancy, and output tolerance rr, let s=r/(2κ)s=r/(2\kappa), where κ=1\kappa=1 for metric outputs, and define

p∗=max⁡cPr⁡[f(x+ϵ)∈B(c,s)],p_* = \max_c \Pr[f(x+\epsilon)\in B(c,s)],

pCS=Pr⁡[f(x+ϵ)∈B(gCS(x),s)].p_{\mathrm{CS}} = \Pr[f(x+\epsilon)\in B(g_{\mathrm{CS}}(x),s)]. By construction, p∗≥pCSp_*\ge p_{\mathrm{CS}}. Therefore, when pCS>1/2p_{\mathrm{CS}}>1/2, the corresponding population majority certificates satisfy

=\sigma\Phi^{-1}(p_*) \ge \sigma\Phi^{-1}(p_{\mathrm{CS}}) =R_{\mathrm{CS}}^{\mathrm{inv}}.$$
AuthorsOpenReview ↗

Response to Reviewer Comments (Part 3)

The inequality is strict whenever p∗>pCSp_*>p_{\mathrm{CS}}. Thus, ContRS maximizes this particular population majority certificate at the fixed neighborhood radius, while inversion of center smoothing preserves the original center. The certification benefit comes from optimizing the probability mass entering the shared Gaussian formula.

In summary, inversion answers how to certify an existing smoothed prediction at a prescribed tolerance. However, inversion does not optimize prediction aggregation or center selection, and the resulting bounds may remain conservative. ContRS instead sets the neighborhood radius according to the prescribed tolerance and selects the center whose neighborhood has the largest probability mass to improve certificate tightness.

[2] Kumar, A., Goldstein, T. Center Smoothing: Certified Robustness for Networks with Structured Outputs. NeurIPS, 2021.

Q4: Does ContRS still outperform safe region when the target probability of the baseline is tuned rather than fixed at p=0.55?
Re: Thank you for this question. Our choice of p=0.55p=0.55 follows the original setting used by safe region in its comparison with center smoothing. To examine sensitivity to this choice, we additionally evaluate a higher target probability, p=0.7p=0.7, across all three tasks, using σ∈{0.20,0.25,0.30,0.35,0.40}\sigma\in\lbrace 0.20,0.25,0.30,0.35,0.40\rbrace. For each input and noise setting, let sSRs_{\mathrm{SR}} denote the radius of the output ball selected by safe region at p=0.7p=0.7. We set the ContRS output tolerance to r=2κsSRr=2\kappa s_{\mathrm{SR}}, so its neighborhood radius r/(2κ)r/(2\kappa) equals sSRs_{\mathrm{SR}}. We report certified input radii and smoothing errors. Larger radii and smaller smoothing errors are better.

Comparison of certified input radii between ContRS and safe region at matched output neighborhood scales.

TaskMethodσ=0.20\sigma=0.20σ=0.25\sigma=0.25σ=0.30\sigma=0.30σ=0.35\sigma=0.35σ=0.40\sigma=0.40
Face Detection (Jaccard Distance)safe region0.10810.14860.24570.35220.5066
ContRS0.23280.37810.68320.74960.6498
Image Reconstruction (Perceptual Distance)safe region0.09770.12200.14670.17150.1959
ContRS0.18850.23600.28060.32830.3698
Image-Text Retrieval (Cosine Distance)safe region0.09700.12210.14620.17100.1950
ContRS0.17700.21920.26640.30910.3352

Comparison of smoothing errors between ContRS and safe region at matched output neighborhood scales.

TaskMethodσ=0.20\sigma=0.20σ=0.25\sigma=0.25σ=0.30\sigma=0.30σ=0.35\sigma=0.35σ=0.40\sigma=0.40
Face Detection (Jaccard Distance)safe region0.09720.12090.16250.24030.3486
ContRS0.08520.10010.12200.13650.1706
Image Reconstruction (Perceptual Distance)safe region0.00070.00130.00210.00310.0042
ContRS0.00070.00130.00210.00300.0041
Image-Text Retrieval (Cosine Distance)safe region0.19330.24690.28920.32310.3517
ContRS0.19200.24600.28920.32260.3516

In general, ContRS achieves larger certified input radii in all 15 evaluated task and noise configurations. Its smoothing error is lower in 11 configurations and equal at the reported precision in the remaining four. These results show that the observed certification benefits persist across three tasks.

Reviewer_R4GwOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ContRS, a framework for extending randomized smoothing to continuous-output prediction tasks. The method selects an output neighborhood containing the largest probability mass under Gaussian input perturbations and derives a certified input radius for a prescribed output tolerance. Experiments evaluate the approach on face detection, image reconstruction, and image–text retrieval.

Critical strengths
  1. The paper addresses a relevant problem: defining and certifying robustness for continuous valued predictions.

  2. The central formulation is intuitive and provides a clear connection between acceptable output variation and certified input robustness. It also recovers classical classification smoothing (for the classification over discrete labels) as a special case.

  3. The evaluation across different output spaces is a strength, and the comparisons with existing methods show encouraging results.

Critical weaknesses

I do not believe the paper has major "critical weaknesses" but it does have some aspects which can be strengthened. Here are a few:

  1. The contribution relative to center smoothing and safe-region certification could be articulated more sharply, particularly regarding differences in constructing the predictor, selecting the output-region, and the guarantees being compared.

  2. The relationship between the ideal smoothed predictor and the finite-sample implementation also can be explained in a more clear fashion. Readers should be able to understand precisely which predictions the certificate applies to and what additional uncertainty is introduced by approximating the selected center.

  3. Finally, although the experiments cover several tasks, the evaluation uses only 50 test inputs per task. Larger test sets and uncertainty estimates would strengthen confidence in the reported improvements.

Questions for authors
  1. Could the authors clarify how the theoretical guarantee relates to the predictions returned by the finite-sample algorithm? -- taken care of by authors in the response.

  2. Do the improvements over center smoothing remain when both methods account for approximation error and certify the same notion of prediction stability? -- taken care of by authors in the response.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Comments (Part 1)

Thank you for your positive assessment of our work. We have provided detailed, point-by-point responses to the weaknesses and questions you raised, and we hope these address your concerns.

Q1: The contribution relative to center smoothing and safe-region certification could be articulated more sharply.
Re:Thank you for this suggestion. We clarify the relationship between ContRS, center smoothing, and safe region along three dimensions: smoothed prediction construction, output region selection, and certification guarantees. The central idea of ContRS is to incorporate the prescribed output tolerance into prediction aggregation and region selection, and to establish the corresponding stability and finite-sample certification guarantees.

AspectCenter smoothingsafe regionContRS (ours)
PredictionCenter of the smallest ball with mass ≥ 1/2Binary region membershipCenter with the largest neighborhood mass
Region selectionFixed center; radius set by input budgetRegion chosen for certificationRadius set by output tolerance; center maximizes mass
GuaranteeOutput deviation bound for a given input radiusMajority mass remains in the selected regionInput radius certifying the prescribed output tolerance

Relative to center smoothing, the distinction lies in the objective for center selection. Center smoothing finds the smallest output ball containing at least half of the probability mass, but its center need not capture the largest probability mass at the neighborhood radius associated with the prescribed tolerance. Consequently, the neighborhood used for certification may contain less probability mass, potentially limiting the certified input radius. ContRS instead fixes the neighborhood radius according to the output tolerance and selects the center that captures the most probability mass. Center selection is directly aligned with the certification objective for the prescribed tolerance.

Relative to safe region, the distinction lies in the output region selection and the certification objective. Safe region first selects an output region and then certifies that, under input perturbations, the probability of noisy model outputs falling within that region remains above 1/21/2. Its main guarantee concerns the probability mass within the selected region, rather than directly bounding the discrepancy between the original and perturbed smoothed predictions. In contrast, ContRS starts from a prescribed output tolerance rr, sets the neighborhood radius to s=r/(2κ)s=r/(2\kappa), and selects the center whose neighborhood captures the largest probability mass as the smoothed prediction.

When the majority condition holds, ContRS uses Gaussian smoothing to ensure that the selected neighborhood retains majority probability mass within the certified input radius. Combining this guarantee with the neighborhood maximization property of the perturbed smoothed prediction and the geometric properties of the output distance, ContRS further establishes

d(gs(x+δ),gs(x))≤r.d\bigl(g_s(x+\delta),g_s(x)\bigr)\leq r.

Thus, ContRS unifies output region selection, smoothed prediction construction, and certification of prediction changes, extending beyond a probability guarantee for a fixed output region.

Q2: The relationship between the ideal smoothed predictor and the finite-sample implementation also can be explained in a more clear fashion.
Re: Thank you for this suggestion. We will clarify three aspects of this relationship: what the implementation approximates, why its certificate remains valid, and what the guarantee covers.

1. What the implementation approximates.

The implementation uses finite samples to estimate the ideal prediction gs(x)g_s(x). The ideal predictor selects the center whose neighborhood contains the largest probability mass. Algorithm 1 makes this selection using samples: it draws noisy outputs as candidate centers, uses independent samples to count how many outputs fall within each candidate’s neighborhood, and returns the candidate with the largest count as z^x\hat z_x. Thus, z^x\hat z_x is the prediction selected from finite observations and need not equal the ideal prediction gs(x)g_s(x).

AuthorsOpenReview ↗

Response to Reviewer Comments (Part 2)

2. Why certification remains valid.

After selecting z^x\hat z_x, Algorithm 1 uses independent certification samples to compute a one-sided Clopper–Pearson lower bound p^L\hat p_L on its neighborhood probability mass. These samples are independent of candidate generation and center selection, ensuring that the confidence bound remains valid after selection. The ideal center gs(x)g_s(x) covers at least as much probability mass as the selected center. Hence, with confidence at least 1−α1-\alpha,

\ge \Pr[f(x+\epsilon)\in B(\hat z_x,s)] \ge \hat p_L.$$ Thus, the probability bound obtained from finite samples also provides a valid lower bound for the ideal center’s neighborhood mass. This is why certification does not require the sampled center to equal the ideal center. **3. What the guarantee covers.** When $\hat p_L>1/2$, Algorithm 1 returns $\hat R=\sigma\Phi^{-1}(\hat p_L)$. Under metric or symmetric relaxed triangle assumptions, Corollary 4.2 guarantees, with confidence at least $1-\alpha$, that the ideal prediction after perturbation remains within $r$ of both the original ideal prediction $g_s(x)$ and the sampled center $\hat z_x$ returned by the implementation: $$d(g_s(x+\delta),g_s(x))\le r, \qquad d(g_s(x+\delta),\hat z_x)\le r$$ for every $\|\delta\|_2<\hat R$. This connects the finite-sample computation to the ideal predictor without requiring the sampled center to equal the ideal center. The guarantee concerns $g_s(x+\delta)$; it does not directly certify a newly sampled center at the perturbed input. **Q3: Readers should be able to understand precisely which predictions the certificate applies to and what additional uncertainty is introduced by approximating the selected center.** **Re:** Thank you for the suggestion. We have further clarified the predictions covered by the certificate and the uncertainty introduced by selecting the center from finite samples. **First, the predictions covered by the certificate.** ContRS defines the ideal smoothed predictor $g_s(x)$ by maximizing population neighborhood probability mass, whereas Algorithm 1 selects a center $\hat z_x$ from finite samples. Corollary 4.2 and its extension provide valid certification without requiring these centers to coincide. Specifically, under the relevant geometric conditions, with confidence at least $1-\alpha$, the following bounds hold simultaneously for every input perturbation satisfying $\|\delta\|_2<\hat R$: $$d(g_s(x+\delta),g_s(x))\leq r,$$ $$d(g_s(x+\delta),\hat z_x)\leq r.$$ Thus, the certificate bounds the change in the ideal smoothed prediction under input perturbations and explicitly relates the perturbed ideal prediction to the center returned by the finite-sample implementation. **Second, the uncertainty introduced by center approximation.** Finite-sample center selection may return a center whose neighborhood probability mass is below the population optimum, potentially making the certified radius more conservative. However, this approximation does not require an additional failure probability in the certification guarantee. Algorithm 1 uses independent certification samples and the Clopper–Pearson method to compute a lower confidence bound $\hat p_L$ on the probability mass of the selected neighborhood. Since the population-optimal center captures at least as much probability mass as the sampled center, the following holds with confidence at least $1-\alpha$: $$\Pr\left[f(x+\epsilon)\in B(g_s(x),s)\right] \geq \Pr\left[f(x+\epsilon)\in B(\hat z_x,s)\right] \geq \hat p_L.$$ Therefore, the certification guarantees above remain valid without exact global optimization of the center, an additional approximation error term, or an additional failure probability allocated to center recovery. **Q4: Larger test sets and uncertainty estimates would strengthen confidence in the reported improvements.** **Re**: Thank you for this suggestion. We enlarg the test set from 50 to 100 inputs per task. We compare with center smoothing and safe region. Smaller is better for the certified output bound and the smoothing error, larger is better for the certified input radius, and the best value in each column is in bold. The certified output bound of ContRS equals its output tolerance r. The 95% confidence interval (95% CI) is that of the paired difference in smoothing error between ContRS and the baseline. ### Face Detection (Jaccard Distance) **Table 1: Certified output bounds for center smoothing and ContRS.** | output tolerance r | 0.08 | 0.10 | 0.12 | 0.14 | 0.16 | |---|---|---|---|---|---| | center smoothing | 0.1193 | 0.1601 | 0.2042 | 0.2473 | 0.2920 | | **ContRS** | **0.08** | **0.10** | **0.12** | **0.14** | **0.16** |
AuthorsOpenReview ↗

Response to Reviewer Comments (Part 3)

Table 2: Smoothing error for center smoothing and ContRS.

output tolerance r0.080.100.120.140.16
center smoothing0.04830.05330.05470.05580.0556
ContRS0.04770.05220.05450.05480.0556
Difference (ContRS − center smoothing) [95% CI]−0.0006 [−0.0019, +0.0007]−0.0011 [−0.0024, +0.0003]−0.0003 [−0.0020, +0.0014]−0.0010 [−0.0017, +0.0002]+0.0000 [−0.0016, +0.0012]

Table 3 : Certified input radii for safe region and ContRS.

noise level σ0.10.20.30.40.5
safe region0.00860.02990.06470.33440.0431
ContRS0.03280.16820.34210.62660.2259

Table 4: Smoothing errors for safe region and ContRS.

noise level σ\sigma0.10.20.30.40.5
safe region0.05680.09720.14100.30260.1468
ContRS0.05680.08050.11100.15950.1458
Difference (ContRS − safe region) [95% CI]+0.0000 [−0.0027, +0.0027]−0.0167 [−0.0419, +0.0084]−0.0300 [−0.0619, +0.0019]−0.1431 [−0.2060, −0.0803]−0.0009 [−0.0036, +0.0017]

Image Reconstruction (Perceptual Distance)

Table 5: Certified output bounds for center smoothing and ContRS.

output tolerance rr0.0120.0140.0160.0180.020
center smoothing0.04000.04400.04910.05480.0613
ContRS (= rr)0.0120.0140.0160.0180.020

Table 6: Smoothing errors for center smoothing and ContRS.

output tolerance rr0.0120.0140.0160.0180.020
center smoothing0.00160.00160.00160.00160.0016
ContRS0.00160.00160.00160.00160.0016
Difference (ContRS − center smoothing) [95% CI]+0.0000 [−0.0001, +0.0001]+0.0000 [−0.0000, +0.0001]−0.0000 [−0.0001, +0.0001]+0.0000 [−0.0000, +0.0001]+0.0000 [−0.0000, +0.0001]

Table 7: Certified input radii for safe region and ContRS.

noise level σ\sigma0.10.20.30.40.5
safe region0.00910.01820.02760.03820.0460
ContRS0.05610.10430.17130.21660.2922

Table 8: Smoothing errors for safe region and ContRS.

noise level σ\sigma0.10.20.30.40.5
safe region0.00020.00090.00240.00500.0085
ContRS0.00020.00090.00240.00490.0083
Difference (ContRS − safe region) [95% CI]−0.0000 [−0.0000, +0.0000]−0.0000 [−0.0001, −0.0000]+0.0000 [−0.0001, +0.0001]−0.0001 [−0.0003, +0.0001]−0.0002 [−0.0006, +0.0002]

Image-Text Retrieval (Cosine Distance)

Table 9: Certified output bounds for center smoothing and ContRS.

output tolerance rr0.0800.0880.1000.1080.120
center smoothing0.25800.27220.29820.31960.3563
ContRS (= rr)0.0800.0880.1000.1080.120

Table 10: Smoothing errors for center smoothing and ContRS.

output tolerance rr0.0800.0880.1000.1080.120
center smoothing0.23370.23370.23370.23370.2337
ContRS0.23320.23220.23320.23370.2336
Difference (ContRS − center smoothing) [95% CI]−0.0005 [−0.0021, +0.0011]−0.0015 [−0.0029, −0.0001]−0.0006 [−0.0018, +0.0007]−0.0000 [−0.0014, +0.0013]−0.0002 [−0.0015, +0.0011]

Table 11: Certified input radii for safe region and ContRS.

noise level σ\sigma0.10.20.30.40.5
safe region0.00920.01790.02760.03600.0442
ContRS0.05970.09590.19140.22100.2093

Table 12: Smoothing errors for safe region and ContRS.

noise level σ\sigma0.10.20.30.40.5
safe region0.09210.18740.26990.32730.3745
ContRS0.09190.18650.26980.32700.3739
Difference (ContRS − safe region) [95% CI]−0.0002 [−0.0011, +0.0008]−0.0009 [−0.0022, +0.0003]−0.0000 [−0.0012, +0.0011]−0.0003 [−0.0017, +0.0012]−0.0006 [−0.0018, +0.0006]

As shown in these tables, on all three tasks and at every output tolerance, ContRS achieves a much tighter certified output bound than center smoothing. It also achieves a much larger certified input radius than safe region at every noise level, with a lower or comparable smoothing error.

Reviewer_1i9XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces ContRS, an approach that extends randomized smoothing to continuous predictions.

Critical strengths

The paper addresses a limitation of conventional randomized smoothing by extending robustness certification.

Critical weaknesses

W1. Motivation of the paper needs to be clearer. In addition, motivation-related figures are needed to demonstrate what previous studies do not address and what this paper aims to address. W2. To demonstrate the effectiveness and generalizability of their method, the authors should evaluate their approach on medical benchmarks e.g., MIMIC-CXR etc. W3. The manuscript lacks sufficient performance comparisons with recent baselines e.g., ARS [1], DRS [2], which are needed to assess the approach effectiveness and generalization. W4. Further evaluation using additional architectures would strengthen the evidence that the proposed work can generalize across diverse base models.
W5. The empirical robustness evaluation should be strengthened with adaptive attacks e.g., BPDA, BPDA-EOT, PGD-EOT, along with AutoAttack, C&W, and recent diffusion attacks. These evaluation would help establish that the proposed work effective ones.

[1] Adaptive Randomized Smoothing: Certified Adversarial Robustness for Multi-Step Defences, NeurIPS, 2024

[2] Mitigating the Curse of Dimensionality for Certified Robustness via Dual Randomized Smoothing, ICLR 2024

Questions for authors

See weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer Comments (Part 1)

Thank you for taking the time to review our paper and share your feedback. We provide detailed responses to each of your comments and questions below. We have also revised Introduction in the updated PDF to explain our work more clearly. We hope these explanations and revisions help address your concerns.

Q1: Motivation of the paper needs to be clearer.
Re: Thank you for this suggestion. We will further clarify the motivation of our paper from the following three perspectives.

First, basic objective. Our goal is to provide a theoretical generalization of classical randomized smoothing to continuous outputs. Classical randomized smoothing aggregates class probabilities under noisy inputs to certify label invariance. For continuous outputs, model robustness typically allow some prediction changes and require these changes to remain within an acceptable range. We therefore aim to retain the certification mechanism based on probability aggregation and comparison while extending it to continuous outputs: given a bound on output variation prescribed by some applications, we determine a certified input radius within which that requirement is satisfied.

Second, existing limitations. As discussed in our responses to Q1 and Q3 from Reviewer QyJY, extensions of randomized smoothing can provide certified input radii for continuous outputs. However, their output representations, aggregation criteria, and certification targets still limit cross-task applicability and certificate tightness. We therefore need a more effective mapping from a prescribed scale of output variation to the certified input radius, reducing adaptation costs across tasks and enlarging the certifiable input region for a given output tolerance.

Third, core contributions. We incorporate the output tolerance directly into the joint construction of the smoothed prediction and certification conditions. ContRS extends the single-label events in classical random smoothing to events defined by output neighborhoods. It sets the neighborhood radius according to the prescribed output tolerance, selects the center whose neighborhood has the largest probability mass as the smoothed prediction, and compares this mass with those of competing neighborhoods centered beyond the allowed tolerance to derive the certified input radius directly. At a fixed neighborhood radius, greater neighborhood probability mass yields a larger radius under the corresponding majority certificate, directly aligning center selection with the certification objective. We further establish theoretical guarantees based on comparisons of competing neighborhood probabilities and, under the relevant geometric conditions, finite-sample certificates that do not require exact recovery of a globally optimal center. When output neighborhoods reduce to single-label events, the framework recovers classical randomized smoothing.

Q2: Motivation-related figures are needed to demonstrate what previous studies do not address and what this paper aims to address.
Re: Thank you for the suggestion. We have added a motivation figure to Introduction. As discussed above, the core motivation behind ContRS is to incorporate the prescribed output tolerance directly into the construction of the smoothed prediction and certification conditions. ContRS sets the neighborhood radius according to this tolerance, selects a prediction center by aggregating the probability mass of nearby outputs, and then derives the corresponding certified input radius. This allows different tasks to use a unified certification framework with their respective output discrepancy functions, reducing the need for task-specific adaptation. Moreover, for a fixed neighborhood radius, when the majority condition holds, selecting a center that captures more probability mass yields a larger certified input radius for the same output tolerance.

In addition, Table 1 compares representative existing methods along three dimensions: output settings, certified properties, and certification directions. It clarifies the certification objective of ContRS: deriving a certified input radius from a prescribed output tolerance to directly bound the discrepancy between the original and perturbed smoothed predictions. We hope that the added motivation figure and Table 1 help readers better understand the motivation of our work and its relationship to existing methods.

AuthorsOpenReview ↗

Response to Reviewer Comments (Part 2)

Q3: To demonstrate the effectiveness and generalizability of their method, the authors should evaluate their approach on medical benchmarks e.g., MIMIC-CXR etc.
Re: Thank you for the suggestion. Access to MIMIC-CXR requires credentialing and a signed data use agreement. We therefore conducted the additional medical imaging evaluation on BCCD [1], a publicly available medical dataset. We use YOLOv8m as the base detector. Predictions are represented as bounding boxes, and output variation is measured using Jaccard distance. For center smoothing, we compare certified output bounds at matched input radii: for each input and output tolerance, center smoothing uses the input radius certified by ContRS. For safe region, we compare certified input radii at matched output neighborhood scales. Smoothing error measures the Jaccard distance between the base and smoothed predictions.

Following the setting of Appendix J, We adjust the noise range to the base detector. We use σ=0.02\sigma=0.02 for the comparison with center smoothing and σ∈{0.01,0.015,0.02,0.025,0.03}\sigma\in\lbrace 0.01,0.015,0.02,0.025,0.03\rbrace for the comparison with safe region.

Table 1: Certified output bounds at matched input radii.

output tolerance rr0.0160.0170.0180.0190.020
center smoothing0.08870.09170.09450.09770.1020
ContRS0.0160.0170.0180.0190.020

Table 2: Smoothing error for the comparison with center smoothing.

output tolerance rr0.0160.0170.0180.0190.020
center smoothing0.04940.04940.04940.04940.0494
ContRS0.04580.04570.04570.04590.0459

Table 3: Certified input radii at matched output neighborhood scales.

noise level σ\sigma0.010.0150.020.0250.03
safe region0.00090.00130.00170.00850.0293
ContRS0.00590.00920.02180.06700.0550

Table 4: Smoothing error for the comparison with safe region.

noise level σ\sigma0.010.0150.020.0250.03
safe region0.02280.03650.04820.14540.3888
ContRS0.02250.03570.04640.11170.3357

As shown in above tables, ContRS provides tighter certified output bounds than center smoothing at matched input radii and larger certified input radii than safe region at matched output neighborhood scales.

[1] Gan, L., Li, X., Wang, X. A Curated and Re-annotated Peripheral Blood Cell Dataset Integrating Four Public Resources. Scientific Data, 2025.

Q4: The manuscript lacks sufficient performance comparisons with recent baselines e.g., ARS, DRS, which are needed to assess the approach effectiveness and generalization.
Re: Thank you for the suggestion. The certification guarantees in ARS and DRS primarily ensure that classification labels remain unchanged under input perturbations, whereas ContRS certifies that changes in continuous outputs remain within a prescribed tolerance. Thus, these methods are not directly comparable baselines in our continuous-output certification setting. To strengthen the evaluation, we have added randomized α-smoothing [2] as a related baseline that studies robustness certification for regression.

For each task, we prescribe the same output tolerance rr for both methods and compare their certified input radii and smoothing errors. The values of rr follow our paper, and the baseline uses its original hyperparameters. For image reconstruction, we use pixel-wise ℓ∞\ell_\infty distance because the baseline’s coordinate-wise certificate does not provide a perceptual-distance guarantee. Larger certified input radii and smaller smoothing errors indicate better performance.

Face Detection (Jaccard Distance)

Table 5: Certified input radius at the same output tolerance.

output tolerance rr0.040.060.080.100.12
randomized α-smoothing1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}0.00030.00190.0036
ContRS0.01180.03220.05460.08590.1233

Table 6: Smoothing error at the same output tolerance.

output tolerance r0.040.060.080.100.12
randomized α-smoothing0.05050.05050.05050.05050.0505
ContRS0.04960.05010.05020.05040.0504

Image-Text Retrieval (Cosine Distance)

Table 7: Certified input radius at the same output tolerance.

output tolerance rr0.0800.0880.1000.1080.120
randomized α-smoothing1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}
ContRS0.07960.11000.15550.19030.2425
AuthorsOpenReview ↗

Response to Reviewer Comments (Part 3)

Table 8: Smoothing error at the same output tolerance.

output tolerance r0.0800.0880.1000.1080.120
randomized α-smoothing0.24220.24220.24220.24220.2422
ContRS0.23510.23580.23680.23750.2386

Image Reconstruction (Pixel-Wise ℓ∞\ell_\infty Distance)

Table 9: Certified input radii at the same output tolerance.

output tolerance rr0.0120.0140.0160.0180.0200.022
randomized α-smoothing1×10−61 \times 10^{-6}1×10−61 \times 10^{-6}1×10−61 \times 10^{-6}1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}1×10−51 \times 10^{-5}
ContRS0.19020.25090.31160.37240.43310.4604

Table 10: Smoothing error at the same output tolerance.

output tolerance r0.0120.0140.0160.0180.0200.022
randomized α-smoothing0.00110.00110.00110.00110.00110.0011
ContRS0.00080.00090.00090.00100.00100.0011

As shown in above results, at every output tolerance, ContRS achieves a larger certified input radius than randomized α-smoothing, with a comparable or lower smoothing error. Randomized α-smoothing certifies almost no positive input radius on all three tasks, because every output coordinate must simultaneously satisfy its tolerance interval.

[2] Rekavandi, A. M., Farokhi, F., Ohrimenko, O., Rubinstein, B. I. P. Certified Adversarial Robustness via Randomized α-Smoothing for Regression Models. NeurIPS, 2024.

Q5: Further evaluation using additional architectures would strengthen the evidence that the proposed work can generalize across diverse base models. Re: Thank you for the suggestion. We have expanded the evaluation by using an additional publicly available base model for each task. These experiments examine whether the certification benefits of ContRS persist under changes in different architectures.

For face detection, we replace MTCNN with RetinaFace-R50 [3], introducing a single-stage detector in place of the original cascaded architecture. For image reconstruction, we evaluate an additional convolutional VAE [4] trained on CIFAR-10, using the publicly available checkpoint. For image-text retrieval, we replace CLIP with SigLIP 2 [5], which uses a different pretraining strategy. The concrete results are listed as follows.

Face detection with RetinaFace-R50

Table 11: Comparison with center smoothing at matched input radii.

MetricMethodr=0.14r=0.14r=0.15r=0.15r=0.16r=0.16r=0.17r=0.17r=0.18r=0.18
certified output boundcenter smoothing0.74700.75680.76990.79980.8273
certified output boundContRS0.140.150.160.170.18
smoothing errorcenter smoothing0.28880.28880.28880.29560.3022
smoothing errorContRS0.19310.19320.19360.19810.2019

Table 12: Comparison with safe region at matched output neighborhood scales.

MetricMethodσ=0.1\sigma=0.1σ=0.2\sigma=0.2σ=0.3\sigma=0.3σ=0.4\sigma=0.4σ=0.5\sigma=0.5
certified input radiussafe region0.00860.01710.04580.06060.3310
certified input radiusContRS0.03780.08670.38470.49461.0399
smoothing errorsafe region0.04110.06840.10890.14380.2885
smoothing errorContRS0.04090.06820.09460.13240.2143

Image reconstruction with convolutional VAE

Table 13: Comparison with center smoothing at matched input radii.

MetricMethodr=0.012r=0.012r=0.014r=0.014r=0.016r=0.016r=0.018r=0.018r=0.020r=0.020
certified output boundcenter smoothing0.03680.04000.04430.04850.0524
certified output boundContRS0.0120.0140.0160.0180.020
smoothing errorcenter smoothing0.00260.00270.00270.00280.0028
smoothing errorContRS0.00260.00270.00270.00270.0028

Table 14: Comparison with safe region at matched output neighborhood scales.

MetricMethodσ=0.1\sigma=0.1σ=0.2\sigma=0.2σ=0.3\sigma=0.3σ=0.4\sigma=0.4σ=0.5\sigma=0.5
certified input radiussafe region0.00900.01800.02650.03740.0448
certified input radiusContRS0.05540.11600.17920.24140.2982
smoothing errorsafe region0.00010.00080.00270.00610.0120
smoothing errorContRS0.00010.00080.00260.00600.0118

Comments on the paper

1 comment

Dear Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs,

We sincerely appreciate the time and effort you have devoted to reviewing our manuscript. We are encouraged by the positive feedback on our work and grateful for the constructive suggestions for improvement. ContRS extends randomized smoothing to continuous outputs by incorporating the prescribed output tolerance into prediction aggregation and certification. We establish population guarantees for prediction changes and, under suitable geometric conditions, valid finite-sample certificates.

In response to the reviews, we have provided detailed, point-by-point replies and expanded our evaluation to include medical imaging, comparisons with new baselines, additional base models, adaptive and diffusion attacks, larger test sets with confidence intervals, and comparisons with safe region at an additional target probability. We have also clarified the motivation and theoretical contributions, distinctions from existing smoothing methods, and the relationship between the ideal smoothed predictor and its finite-sample implementation, including the precise scope of certification guarantees.

We hope these additions and clarifications address your concerns. We would be happy to answer any further questions before the discussion deadline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.