Stability Guarantees for Feature Attributions with Multiplicative Smoothing

Explanation methods for machine learning models tend not to provide any formal guarantees and may not reflect the underlying decision-making process. In this work, we analyze stability as a property for reliable feature attribution methods. We prove that relaxed variants of stability are guaranteed if the model is sufficiently Lipschitz with respect to the masking of features. We develop a smoothing method called Multiplicative Smoothing (MuS) to achieve such a model. We show that MuS overcomes the theoretical limitations of standard smoothing techniques and can be integrated with any classifier and feature attribution method. We evaluate MuS on vision and language models with various feature attribution methods, such as LIME and SHAP, and demonstrate that MuS endows feature attributions with non-trivial stability guarantees.

Paper

References (64)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer WB7f6/10 · confidence 3/52023-07-01

Summary

The paper is about the stability of explanations in the feature attribution setting for image classification tasks. They illustrate that in certain settings, swapping/removing a pixel of an explanation completely changes the classifiers prediction, which is undesired. They define what a "stable" explanation is formally and introduce Multiplicative Smoothing as a method to obtain these stable explanations. In principle it's multiplying the image with feature attribution masks such that the classifier becomes smooth in the sense that it doesn't suddenly (non-smoothly) change it's prediction from for example cat to goldfish. Experiments demonstrate the efficacy of MuS.

Strengths

1. It's clear what the paper is trying to achieve 2. Solution proposal is simple and thoroughly explained with good theoretical justification.

Weaknesses

1. It's a bit unclear what exactly smoothness refers to, currently it seems to refer to no discontinuous jumps when removing a feature attribution pixel in classification output? 2. The experiments are a bit unintuitive, needs clarification what's being tested here and why.

Questions

1. What's smoothness exactly, never got a good intuition what this exactly means. 2. Don't understand what the experiments do, please explain how and why these experiments demonstrate that MuS is working as intended. 3. Is it correct to understand that the paper is in some sense looking for the "minimal" set up features that explains a class that won't change with some perturbation ? ----- Post rebuttal: Authors did a good job explaining what's going on, left some feedback for exposition, raised to 6, it's an interesting paper.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

N/A

Reviewer WjBG6/10 · confidence 4/52023-07-02

Summary

This paper aims to make the classifier robust to feature removal and addition. The authors find that adding a patch to the mask obtained from explanation method may cause the classifier to make substantially different prediction. They introduce the notion of incremental stability and decremental stability to measure classifier's stability w.r.t mask change. They introduce MuS to any base classifier and show that MuS provides stability guarantee for this classifier.

Strengths

1. The notion of incremental stability and decremental stability is novel. 2. They motivate multiplicative masking instead of the widely-adopted additive masking by pointing out the inconsistency involved in additive masking. 3. Certified stability guarantee is obtained. 4. A sample-efficient algorithm is introduced to reduce the sample complexity to sample from Bernoulli distribution.

Weaknesses

1. Although masks can be sampled efficiently, to get the final prediction, a lot of inference steps are involved. This might be time-consuming. The authors do not talk about how this could be tackled. 2. MuS seems to be similar with Randomized Smoothing with Bernoulli distribution. By averaging predictions under several masks, the classifier use information from the whole image to make prediction. It is not too surprising that the classifier is more robust to mask change.

Questions

1. Are robust models also sensitive to mask removal or addition? 2. There are many works discussing robust attributions (e.g., [1], [2]) and establishing the connection between the robustness of explanations and the robustness of black-box models (e.g., [3], [4]). What's the relationship between these works and this paper? [1] Dombrowski et al., Explanations can be manipulated and geometry is to blame, NeurIPS 2019 [2] Wang et al., Smoothed Geometry for Robust Attribution, NeurIPS 2020 [3] Tan et al., Robust Explanation for Free or At the Cost of Faithfulness, ICML 2023 [4] Agarwal et al., Rethinking Stability for Attribution-based Explanations, ICLR 2022

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Listed is Weakness and Questions sections.

Reviewer JJJj6/10 · confidence 2/52023-07-11

Summary

In this paper the authors have introduced a framework to measure the stability of feature attribution methods. They do so by introducing two relaxed notions of stability called incremental stability and descremental stability which check for stability in a neighbourhood of the original feature set. They show that size of these stable neighbourhoods can be measured for lipschitz smooth classification functions and then present a methdology to convert any classified into a Lipschitz smooth classifier using random sampling. Finally they conclude the paper with numerical experiments to illustrate the benefits of their approach.

Strengths

I feel that this paper has several strengths. Understanding the stability of feature attribution methods is a key requirement to evaluate them. This paper presents a novel and efficient approach to understanding these stability metrics which can then be used to compare quality of the features identified by the different methods. The stabiilty metric makes intuitive sense and the theoretical results justify the measurement techniques for them. They also present a way of how this stability metric can be extended to any classifier. Overall, I believe this paper makes a key contribution in this area.

Weaknesses

The metric seems more descriptive than prescriptive. It would be difficult to some how incorporate he metric into the optimization problem so as to guide the network and method towards more stable models and feature attributers.

Questions

I think it is important to provide a discussion about what the impact of Mulitplicative smoothening is on the original classifier. How is the stability metric of the smoothened classifier associated to the original classifier.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Adequately addressed.

Reviewer dPnc7/10 · confidence 4/52023-07-12

Summary

This paper presents a technique for extracting feature attributions that are certifiably stable in the sense that the model's predictions are consistent on supersets of attributed features. As the title suggests, the approach is based on multiplicative smoothing, a novel type of Bernoulli smoothing based on masking, rather than perturbing, input features. Notably, the paper shows how to construct a distribution with coordinate-wise Bernoulli features that requires significantly fewer samples to obtain a certificate.

Strengths

This work introduces a new type of attribution stability that is not addressed by prior work on robust feature attribution, and presents rigorous yet practical techniques for obtaining stable explanations. As the core technique is based on randomized smoothing, it is applicable broadly to different types of models and even different feature attribution methods. Thus, it complements a large body of existing work, and gives concrete, practical ways to improve it with respect to stability. While smoothing-based certification methods are typically very expensive, the structured dependency sampling described in section 3.3 gives a clever way to avoid this, making the approach reasonably inexpensive. The empirical results give a sense of what the guarantees look like on real data, and compare several widely-used attribution methods under the certification regime. The writing is polished and clear on most of the important points. Overall the paper is an enjoyable read.

Weaknesses

The accuracy penalty imposed by multiplicative smoothing is large (a related small point: the horizontal axis between figures 4 and 5 isn't the same, making these results a little harder to line up). It doesn't seem that the models were trained with augmentations that match smoothing noise, or that denoising was used on smoothing samples, so there may be opportunities to improve on this. Without improvements, the results in Figure 5 pull somewhat against the practical significance of the work. The writing is a bit unclear around the sufficiency of Lipschitzness and masking equivalence. A less-careful read of the paper that focuses more on early sections might come away with the impression that most techniques for obtaining Lipschitz models will suffice for the stability guarantees sketched in remark 2.6. The authors might consider moving some of the intuitions from proposition 3.1 into this discussion. The notion of stability at the center of this work is different than what I would have expected, being familiar with some prior work on this topic and reading the abstract. I agree that the model needs to play a central role in characterizing the stability of explanations, but it's less clear that being stable to added features is always an important property. The paper might have wider appeal if the authors are able to justify this, using examples or otherwise (Figure 1 is indeed an example, but doesn't address the question of why the classifier's behavior over the right two images is harmful or undesirable).

Questions

1. Were the models used in figures 4 and 5 trained on appropriately augmented data? 2. Have you attempted to use denoising methods as an intermediate step after sampling from the smoothing distribution? 3. Are there other existing methods for obtaining Lipschitz models, perhaps without smoothing, that could also work with your certificates?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The paper does not have an explicit limitations section, but these points are discussed where it is appropriate.

Reviewer 7LaM7/10 · confidence 2/52023-07-23

Summary

The draft addresses an important problem of feature attribution method selection and formalizes a notion of attribution stability to do this. The approach is based on modifying classifiers to make them Lipschitz wrt to feature masking. Once done, the new classifier has provable radii of stability defined as L1 distances within which the found sparse explanation is sufficient and adding new features will not change the prediction. The approach is tested on vision and NLP tasks with four different explanation methods. Despite some weaknesses the paper is a good contribution to the field of feature attribution explanations.

Strengths

- new measurable definition of explanation stability - a method to convert any classier into a one that permits the stability calculation - theoretical analysis; efficient computation - comprehensive empirical evaluation

Weaknesses

- The need to modify the classifier, before guarantees can be calculated, somewhat diminishes the practical value. In practice, we would be interested in the classifier that actually does the prediction, not a smoothed version of it; as the authors in Sec. 4.1 measured the drop of accuracy from the smoothing to be tens of %s, it could be a no-go for critical classifiers. - The proposed stability is operationalized only for binarized explanations and features as image patches, which throws the classifier into the out-of-distribution regime and, on top of smoothing, further complicates drawing conclusions from the approach. In fact all feature attribution methods compared with proposed stability are continuous and require binarization before they can be used.

Questions

- typo in the last word of line 44 - nitpick: logits are usually defined as pre-softmax values, not probabilities (line 71) - Figures are not readable in black and white

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

No dedicated Limitations section but limitations mentioned in the text

Reviewer Qhau5/10 · confidence 3/52023-07-25

Summary

This paper studies the stability of binary attributions. The attribution is defined to be stable if the prediction does not change when adding additional features. The multiplicative smoothing is proposed to achieve the Lipschitz condition, which is proved to infer the relaxation of stability. Experiments verify that the smoothing technique is useful for stability.

Strengths

1. The paper is well-written and easy to understand. The problem of attribution stability is novel. 2. The proposed multiplicative smoothing is effective under the defined stability setting and can be applied to any models.

Weaknesses

1. For the smoothing technique, my main concern is its application scope, which currently focuses only on binary attribution methods. Although a method is proposed to convert non-binary attributions into binary ones, this conversion process disregards the continuous information in the attribution maps. The potential loss of information during this conversion is significant, especially when dealing with features grouped into 64 patches. 2. The motivation behind this work can be emphasized further. While it is mentioned that explanations can be fragile, more details are necessary to inform readers about the specific circumstances under which explanations may fail. In order for the stabilty of explanations to be meaningful, classifiers must rely solely on attributions, or more features added on attributions, to make correct decisions. However, it remains unclear why practitioners would opt to use attributions for decision-making when they already have access to the original input data. 3. The accuracy drop can also be a weakness of the proposed method. While the decrease is expected, a significant decrease in accuracy may suggest a limited usefulness of the method.

Questions

1. I'm interested in the performance comparison of additive smoothing and multiplicative smoothing on the stability setting.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

See weakness

Reviewer 7LaM2023-08-10

Thank you

Thanks for the detailed reply - keeping the Accept note. (review slightly edited for clarity/typos)

Reviewer JJJj2023-08-11

Response

I thank the authors for the detailed response. I will maintain my "weak accept" review.

Reviewer WjBG2023-08-13

Response

Thanks for the detailed reply! I will raise my confidence score to 4. I have a few follow-up questions and I hope you can give some further explanations. 1. To obtain a prediction, how many inference steps are required? If ~100 inference steps were required, would the smoothing step be very slow? 2. In the experiment results you provide, it seems that with MuS, non-robust model has a higher hitting rate than $L_2 (\epsilon=3), L_\infty (\epsilon=4)$ trained robust models. Do you have any explanation on this result?

Authorsrebuttal2023-08-13

**Q1. Cost of inference steps.** Yes. If $q=100$ samples are used, evaluating the smoothed model would be $\times 100$ slower than the base model. Our method for constructing the distribution allows us to decide the number of samples $q \geq 2$, and we arbitrarily set either $q = 64$ in most of our experiments and sometimes $q = 16$. In developmental testing, we have also found $q = 8$ to perform well (i.e. certified accuracy like Figure 5). Nevertheless, the Reviewer raises an interesting set of additional experiments to run for $q$ selection and we will include them in our revised manuscript. Below is a sample of the certified accuracy numbers (like Figure 5 / Section 2 of Rebuttal Supplementals, there both ran with $q=64$) when we run Vision Transformer at $\lambda = 1/8, 2/8, 3/8, 4/8$ for sample complexities of $q = 4,8, 16, 32, 64, 128$. On a $N = 2000$ sample from ImageNet we obtain the following smoothed classifier accuracies: $\lambda = 1/8$: N/A, 0.662, 0.688, 0.693, 0.695, 0.691 $\lambda = 2/8$: 0.717, 0.743, 0.749, 0.754, 0.752, 0.755 $\lambda = 3/8$: N/A, 0.760, 0.774, 0.777, 0.782, 0.778 $\lambda = 4/8$: 0.776, 0.779, 0.793, 0.790, 0.791, 0.795 Note that some values for $q = 4$ are N/A because this is an insufficient discretization granularity for specifying the corresponding probabilities at $\lambda = 1/8, 3/8$. We observe that in general, increasing the value of $q$ slightly increases the accuracy of the smoothed classifier. This may be because, heuristically, smaller values of $q$ mean that the smoothed classifier is more susceptible to the choice of initial random seed. **Q2. Robust vs non-robust model performance.** We were also surprised by the non-MuS performance of the different models, and we can only speculate as to the cause. One possibility is that some radii used in $L^p$ robustness training were too small to generalize to our setting. Our finding that $L^\infty (\varepsilon = 8)$ did in fact outperform the non-robust case is some preliminary evidence of this.

Reviewer Qhau2023-08-15

Thanks for the author's detailed response, which has addressed most of my concerns. The work's contribution is commendable, especially considering it as a first attempt in explaining stability. Therefore, I will increase my score to 5.

Reviewer WB7f2023-08-16

Re: Rebuttal

Thanks for the clarifications! It's a bit more clear now what you're trying to achieve, it would be appreciated if you could associate the definitions with perhaps more figures or illustrations providing the intuition for stability. Happy to buy in and raise to 6.

Reviewer dPnc2023-08-16

Reply to rebuttal

Thank you for taking the time to write this detailed rebuttal. Your suggestions for addressing the weaknesses mentioned in my original review look reasonable. A few comments on your answers to my questions below. * [Q1] I'm glad to see that training with augmentation improves things! * [Q2] Certainly, additive denoising isn't right. Something based on a diffusion model, along the lines of [1], could be promising. * [Q3] Agreed, and to be clear, I believe that this paper stands on its own either way. I'll maintain my score and support for this paper. [1] Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham, Leslie Rice, Mingjie Sun, J. Zico Kolter. (Certified!!) Adversarial Robustness for Free! https://arxiv.org/abs/2206.10550

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC