Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness

Randomized smoothing is a popular certified defense against adversarial attacks. In its essence, we need to solve a problem of statistical estimation which is usually very time-consuming since we need to perform numerous (usually $10^5$) forward passes of the classifier for every point to be certified. In this paper, we review the statistical estimation problems for randomized smoothing to find out if the computational burden is necessary. In particular, we consider the (standard) task of adversarial robustness where we need to decide if a point is robust at a certain radius or not using as few samples as possible while maintaining statistical guarantees. We present estimation procedures employing confidence sequences enjoying the same statistical guarantees as the standard methods, with the optimal sample complexities for the estimation task and empirically demonstrate their good performance. Additionally, we provide a randomized version of Clopper-Pearson confidence intervals resulting in strictly stronger certificates.

Paper

Similar papers

Peer review

Reviewer xWqy5/10 · confidence 4/52024-07-07

Summary

This paper presents two statistical innovations on top of standard randomized smoothing. The first is employing a randomized Clopper-Pearson interval (instead of deterministic), which marginally increases the certified radius for a particular number of samples. The second involves improving sample efficiency by leveraging confidence sequences, resulting in a roughly 2x speedup over existing methods.

Strengths

I'm torn on this work -- although it seems to be a technically sound improvement on the SOTA, the method is very involved for a relatively small benefit. The paper could also use some better exposition. I'm curious what other reviewers think. 1. This work leverages interesting statistical techniques to improve upon randomized smoothing, both in terms of certified radius and sample efficiency. 2. The theory is quite extensive and seems sound (although I have not checked the proofs in detail). 3. Certified robustness of classifiers is an interesting and topical problem.

Weaknesses

1. The authors should take care to proofread their work, which has a number of grammatical and structural errors. 2. The union-bound and betting-based confidence sequences seem to yield very similar results. I am not sure why both need to be included. 3. The paper flow is somewhat poor and difficult to follow. The authors should revise their manuscript to add better transitions between sections. 4. The proposed approaches include a great deal of complexity for relatively minimal improvement.

Questions

1. Why is the blue line in Figure 1(a) so jagged? 2. I don't understand the purpose of the bounds mentioned on line 189: "In the analysis that follows, we assume that the Bernoulli distribution has a success probability satisfying 0 < c < p < C < 1 for some constants c, C, and thus we can hide the dependence on it into ≍." 3. I have trouble understanding the plots in Figure 2. The text of the paper suggests that epsilon is the width, so why is it plotted from 0 to 1 in the bottom left plot? What exactly are these bounds converging to? I feel that these plots are explaining something important but I'm not sure what the experimental setup is here exactly. 4. The randomized interval in Definition 2.3 only differs from the deterministic interval in the "knife edge" case where the binomial distribution is exactly x. I'd expect the contribution from this term to become negligible as the number of samples grows -- why is this not the case in Figure 2b?

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

The authors do not discuss limitations.

Reviewer NzaR7/10 · confidence 4/52024-07-12

Summary

This paper proposes sample-efficient methods for computing probabilistic robustness certificates for randomized smoothing. The proposed methods replace the standard Clopper-Pearson confidence interval on the classifier’s score, with a confidence sequence, thereby allowing the number of samples to be determined adaptively given a radius of certification $r$ and significance level $\alpha$. Two variants of the method are proposed: one that updates the sequence on a geometric schedule (shown to achieve the asymptotically optimal sample complexity) and another that adopts a betting strategy. The methods are shown to consume 1.5–3 times fewer samples compared to prior work (Horváth et al., 2022).

Strengths

**Originality:** This paper imports methods from statistics/probability theory, which have not previously been applied in the context of randomized smoothing. The methods appear to be an excellent fit for improving the statistical efficiency (and hence computational efficiency) of randomized smoothing. It’s worth noting that the methods have been adapted, in that the bounds are specialized for Bernoulli random variables (rather than generic bounded random variables). **Significance:** The computational cost of randomized smoothing is a barrier to adoption, so it’s great to see work in this direction. The proposed algorithms for adaptively determining the sample size is relatively simple to implement which should encourage adoption. **Quality:** The method is well-motivated. The theory and experiments are generally well-executed, apart from some minor issues outlined below. **Clarity:** The writing is reasonably clear, apart from some issues discussed below.

Weaknesses

**Clarity:** The paper is mostly clear, however some parts could be improved: - Section 3: Definition 3.1 seems out-of-place: it’s not clear that a generic formulation of sequential decision making is needed, if the experiments focus on randomized smoothing. Section 3.1 includes some discussion of the results, but is labeled “related work”. It would be good to add a few more sentences discussing the results - Section 2.2: I find it confusing that symmetric/asymmetric confidence sequences are discussed before confidence sequences are defined (even informally). - Line 189: I found this paragraph confusing – I wonder if it could be explained more concretely (with an example) in an appendix. **Impact:** - The formulation of randomized smoothing in Section 2 covers additive smoothing mechanisms, where the certificate is an $\ell_p$-ball. I suspect the formulation could be generalized to capture non-additive smoothing mechanisms and more general certificate geometries without impacting the validity of the results. - The proposed method is not applicable if one wants to estimate the maximum certified radius at an input. I wonder whether the method could be adapted to estimate the maximum certified radius within some tolerance. **Minor:** - line 24: Provide citation for claim that randomized smoothing “is currently the strongest certification method” - line 62: “realizations **are** lowercase” - line 77: Extraneous closing bracket - line 88: “de-randomized” has not been defined yet. Consider defining earlier or providing a citation. - line 92: Provide citation for claim that the Clopper-Pearson interval is “well known to be conservative” - line 94: In what sense is the confidence interval “optimal”? In terms of coverage? - line 121: “even” → “event” - line 128: Is the case where u() = 0 an upper confidence interval? - Proposition 2.4: Coverage has not been defined for a randomized confidence interval. - line 180: Should $\in$ be $\subseteq$? - line 247: Provide citation for Ville’s inequality - Table 1 caption: Delete “of number of samples” - line 324: “improved them at places” is ambiguous. - line 328: I’m not convinced that we now have a “perfect” understanding of statistical estimation for randomized smoothing. For instance, this paper has not considered the problem of estimating the smoothed classifier’s prediction. It’s possible there may be more sample-efficient ways of estimating the smoothed classifier’s prediction and bounds on the top-2 scores jointly. - “We stress out <some statement>” sounds unusual. It would be more natural to say “We stress <some statement>”.

Questions

- Does the method apply to the more general formulations of randomized smoothing? - Could the method be adapted to estimate the maximum certified radius within some tolerance? This would cover another common use case.

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations ought to be discussed more. For example, the experiments only cover one dataset/model, the proposed method only improves sample complexity if the radius is fixed.

Reviewer mvjJ6/10 · confidence 2/52024-07-12

Summary

They study the task of certified robustness, i.e. they need to decide if a point is robust at a certain radius or not, using as few samples as possible while maintaining statistical guarantees. Their main contribution is utilizing confidence sequences (instead of confidence intervals) that allows them to draw just enough samples to certify robustness of a points which allows them to greatly decrease the number of samples needed. They also show the effectiveness of their approach experimentally. Beyond that, they propose a randomized version of Clopper-Pearson confidence intervals for estimating the class probabilities. A standard component of randomized smoothing procedures is the Clopper-Pearsons confidence interval which is known to be conservative. As a result, the certification procedures underestimate the certified robustness. They provide an optimal confidence interval for binomial random variables that resolves this issue.

Strengths

Certainly one of the main issues of Randomized smoothing methods is that they are not practical due to their computational burden, i.e. a lot of samples need to be drawn to decide robustness at a certain radius. Given that, their result seems interesting.

Weaknesses

Other suggestions: Line 118: hence instead of whence Line 121: event instead of even Contributions paragraph: Perhaps in the first paragraph mention that your results hold for binomial random variables.

Questions

same as weaknesses.

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

same as weaknesses.

Reviewer NzaR2024-08-08

The authors' responses to my two questions have clarified that the proposed methods have broad applicability in randomized smoothing. I have therefore increased my score for "Contribution" and the overall rating. > Yes, our methodology can be directly applied to all[1] randomized smoothing works we are aware of as they have the same estimation subroutine. ... We state this explicitly in the paper. This was not clear to me from reading the paper. In section 2, randomized smoothing is formulated for additive noise and metric balls induced by a norm. Perhaps the authors could include a remark that this more limited formulation is used for ease of exposition, noting that the methods apply more generally as discussed in the response above.

Reviewer xWqy2024-08-10

Thank you to the reviewers for their clarifications. I'm raising my score to a 5, but am still not confident on this paper. Mostly I'm not sure that a 50% reduction in samples is significant, as randomized smoothing is very far from being practical in any sense without an order-of-magnitude conceptual breakthrough. This seems unlikely given that these approaches have been explored for a few years now. While this paper doesn't particularly excite me, perhaps the AC's tastes are different in this regard. Also as a small note: I recommend adding periods after all inline subsections (e.g., after "Related Work" in line 181).

Authorsrebuttal2024-08-10

Thanks a lot! If you think that RS is impractical at this stage, then you can see our paper as a negative result, since our lower bounds prevent significant improvements for the considered task. On the other hand, we provide an adaptive estimation procedure that can be used for virtually any task related to randomized smoothing estimation and we can draw samples one-by-one. Thanks again for your review, we will add the commas.

Reviewer mvjJ2024-08-12

I went through other reviews and responses. Due to the computational cost of RS it is hard to adopt them, and they show an empirical method outperforming SotA by 50% reduction in samples. But it is true that RS is very expensive and it is not clear that 50% reduction in the number of samples is good enough. Furthermore, they show lower bounds that significant further improvements on this task are impossible. I believe the contribution is sufficient enough for acceptance.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC