Red Teaming Deep Neural Networks with Feature Synthesis Tools

Interpretable AI tools are often motivated by the goal of understanding model behavior in out-of-distribution (OOD) contexts. Despite the attention this area of study receives, there are comparatively few cases where these tools have identified previously unknown bugs in models. We argue that this is due, in part, to a common feature of many interpretability methods: they analyze model behavior by using a particular dataset. This only allows for the study of the model in the context of features that the user can sample in advance. To address this, a growing body of research involves interpreting models using \emph{feature synthesis} methods that do not depend on a dataset. In this paper, we benchmark the usefulness of interpretability tools on debugging tasks. Our key insight is that we can implant human-interpretable trojans into models and then evaluate these tools based on whether they can help humans discover them. This is analogous to finding OOD bugs, except the ground truth is known, allowing us to know when an interpretation is correct. We make four contributions. (1) We propose trojan discovery as an evaluation task for interpretability tools and introduce a benchmark with 12 trojans of 3 different types. (2) We demonstrate the difficulty of this benchmark with a preliminary evaluation of 16 state-of-the-art feature attribution/saliency tools. Even under ideal conditions, given direct access to data with the trojan trigger, these methods still often fail to identify bugs. (3) We evaluate 7 feature-synthesis methods on our benchmark. (4) We introduce and evaluate 2 new variants of the best-performing method from the previous evaluation. A website for this paper and its code is at https://benchmarking-interpretability.csail.mit.edu/

Paper

References (71)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer PpyC6/10 · confidence 4/52023-06-27

Summary

This work focuses on uncovering a principled, objective approach to evaluating existing interpretability tools in identifying humanly interpretable attributes in image classification. They approach this via the following synthetic task: 1. A known, humanly interpretable trojan (a special patch, style transfer, a key object) is injected into a model's training pipeline, leading to a false classification by the model when the trojan is present in an image. 2. Different interpretability tools are applied to this model on an uncorrupted image to identify "interpretable features" that would lead to a misclassification. An interpretability tool that performs well should ideally output something about the underlying trojan scheme, since this is a clearly interpretable feature that changes the classification. 3. The quality of outputs from these interpretability tools are compared by seeing which tool's outputs best help humans identify the original trojan scheme from a multiple-choice set of possible options. A key challenge of this task is to find these trojan backdoors with access to the model alone and not to the adversarially corrupted examples. The authors also propose two novel variants on the existing "robust feature-level adversaries" approach of [Casper and Nadeau et.al. Robust feature-level adversaries are interpretability tools] to solve this task. Their main findings/contributions are: 1. The authors propose this trojan identification task as a synthetic benchmark for assessing the usefulness of human-level interpretability tools 2. They claim that naive approaches like saliency maps are not useful, even with access to the adversarial examples, 3. In a human study, robust feature-level adversaries and its variants are shown to be best at helping people identify the backdoor scheme, suggesting that these are the most effective interpretability tools. 4. No technique worked well for identifying style-transfer based, non-local backdoors. This suggests that current interpretability tools might not do a good job of explaining "textural" features.

Strengths

• The paper investigates a useful and important research question in the field of interpretability research: "how do we objectively quantify and measure the quality of interpretable explanations?". • The trojan task proposed by the authors is a clever proxy for measuring human interpretability of explanations. • The authors run an actual human study to measure the usefulness of interpretability tools to people, and their proposed evaluation is a solid measure of whether these tools provide utility. • The authors are upfront about the limitations of their work and the need for diverse techniques other than theirs to create an effectve benchmark for interpretability tools

Weaknesses

• The trojan tasks are fairly simplistic and might not measure the actually important capabilities needed by interpretability tools in the real world. For example, if a tool performs well on this benchmark, it's not clear what that would tell me about how to use it in interpreting my model outputs for a misclassified, nonadversarial image.  The trojan task is a proxy, and I would have loved to see how the "best" models under this proxy can actually be used for a useful interpretability task like identifying why a model is misclassifying certain images. I would have particularly liked to see how their proposed SNAFUE would work in a real interpretability task. • I'm not sure the authors' evidence for the ineffectiveness of saliency maps is well-founded. They use $\ell_1$ error between the saliency map and the location of true patch, which seems like a strange objective. If the patch is small, the $\ell_1$  distance of a perfect saliency map would decrease with patch size, meaning an all-zeros baseline would appear great. If you used a different metric like bounding box intersection you wouldn't get such a strong negative result. *This isn't a huge deal* because this entire section seems not incredibly relevant to the paper, since your main goal is to do this task without access to the adversarial examples. • It took me several reads to make sense of this paper. The writing is not completely clear and the figures took a while to interpret correctly. I would suggest really working on finding a better way to explain the human study task to your readers.

Questions

• On the first few reads of the paper, I thought the goal was to show which tools are best at identifying backdoors in DNNs (the title of the paper further corroborates this interpretation). If that is the case, I would have many questions about why we should care about these artificial tasks rather than, say, a cryptographic planted sequence that can't be inspected or interpreted visually. After several reads, I think what you actually care about is just evaluating how good these "humanly interpretable" tools are, and the trojan task is just a proxy for this. This was not obvious to me and I would suggest a better title and a clearer intro to better emphasize the goal of your paper. • I wasn't sure about the details of your human study. Was it 100 participants per method or 10 per method? It would also be great to see some error bars here.

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

2 fair

Contribution

3 good

Limitations

The authors have a healthy discussion about the limitations of their work. Namely that: 1. They do not cover all possible types of features that may cause model failures.  2. Evaluations are based only on multiple choice questions, which may be sensitive to aspects of survey and experimental design

Reviewer WymJ7/10 · confidence 3/52023-07-03

Summary

This paper studies the performance of different interpretability methods (feature attribution and synthesis) on model debugging, especially for poisoning-based backdoor attacks. The authors also propose to consider this evaluation as a better benchmark of interpretability tools since we know the ground-truth features and It also has more practical applications. For feature attribution methods, they fail to reverse backdoor features in the training dataset even with known backdoored samples. For feature visualization (synthesis) methods, they perform slightly better performance, especially for methods on robust adversarial features. However, the generated features are still hard to understand for humans. Therefore, the authors propose to utilize natural adversarial patches to replace generated adversarial patches.

Strengths

I like this work. This paper tries to answer two important problems of interpretability methods: 1. The Faithfulness of generated feature visualizations: I think there have been two serious problems with feature visualization all along: 1.1 faithfulness: could the generated patterns truly reflect the inner mechanism of DL models? Or just some bias from the optimization method to find them? 1.2 How do humans correctly understand the generated features? are faithful patterns always understandable by humans? 1.3 the natural adversarial patterns satisfy the second point at least, although they also have shortcomings (can not find unseen features). 2. The proper evaluation benchmark: The authors consider the more practical scenario, model debugging, which can more realistically measure the effectiveness of interpretability methods.

Weaknesses

1. About full visualizations of feature attribution with known backdoored samples: I expect that the authors show all visualization results about backdoor reversing. I think the L_1 distance can not fully reflect the poor performance of feature attribution methods. Visualization results are a more proper proxy. I recommend the authors could put visualization in main submission. 2. About attack success rate and clean accuracy of backdoored models: I suggest the authors should give more results of attack success rate and clean accuracy of backdoored models with different triggers. They only mention them on the training dataset on Page 4.

Questions

1. About embedding space of SNAFUE method: do we have better metrics or semantic space to find natural adversarial patches? For me, I think SNAFUE is a variant of the original method based on robust adversarial features, with adding constrained projection. The semantic space and metrics for projection are very important. 2. Could the authors explain why adding natural patches in other class images could mislead DL models and how to further improve this attack performance? It seems that patch attack is not as powerful as PGD attack on L_inf or 2 norm ball. Do you think this is also the reason why "Copy/paste attacks between dissimilar classes are possible but more challenging"? 3. Could you explain more about the 2nd point of Failure Modes for SNAFUE on Page 22? 4. Why visualization of inner neurons is not effective? (In Page 8) Please give more insight into this phenomenon. 5. What are your opinions about the 1st point in Strengths part?

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

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The main limitations are from SNAFUE method. It can not find unseen features. The authors also mentioned it.

Reviewer ChpR4/10 · confidence 2/52023-07-06

Summary

This paper aims to use interpretation tools to identify unknown bugs in out-of-distribution (OOD) models. It highlights the lack of capability to analyze features that cannot be sampled or identified in advance. To resolve this, authors propose to train models that respond to specific triggers and evaluate interpretability tools based on how they can help humans to identify these triggers. Authors propose trojan discovery as an evaluation task and present a benchmark with 12 trojans of 3 types. It assesses the difficulty of the task through the evaluation of 16 feature attribution/saliency tools and 7 feature-synthesis methods. Lastly, authors introduce and evaluate 2 variants of the best-performing methods.

Strengths

I am not an expert in this field and haven’t had any prior experience working in this field, so my evaluations are an educated guess. Quality and significance: The authors conducted extensive experiments benchmarking 12 trojans and 16 feature attribution/saliency tools. The authors also included a comprehensive discussion section with honest limitations, which I highly appreciate. So I think this paper meets the quality and significance standard. Originality: the authors proposed two novel variants of the best performing methods, though the first one “robust feature level adversaries via a generator” is only marginally novel. So I think this paper is on the borderline of the originality standard.

Weaknesses

My main concern for this paper is the poor writing clarity. Understandably, this paper contains a lot of information and thus writing is harder than usual. But for someone new to this field like myself, I find it very hard to follow the big picture and also I find it hard to understand exactly what the authors did. There are also grammar issues here and there, or maybe the grammar is technically correct but the style is unusual, making it hard to read. For example: line 209 “we introduce using …” line 94 the logic of “instead” is not very straight forward. line 79 “[27] who used” I think “who” here is redundant? Lines 56-57: “one which … another which”. Other than grammar issues, the overall writing is just not very clear and easy to follow, and it feels the paper was rushed and not polished. This is the main reason for my rating of 4. I think this paper is not yet ready to be circulated amoung the neurips community.

Questions

See the weakness section.

Rating

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

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

1 poor

Contribution

3 good

Limitations

Yes, the authors have adequately addressed the limitations.

Reviewer EMti6/10 · confidence 3/52023-07-06

Summary

In this work authors study the capabilities of interpretability tools in identifying trojans. Towards this goal authors curate a benchmark dataset and perform both human and automated studies to evaluate various interpretability tools. They discover that their benchmark can be a challenging benchmark for most of these tools and to improve some of these techniques they build upon two previously existing methods.

Strengths

1. The problem they are studying is interesting. 2. The curated benchmark can be useful to some communities.

Weaknesses

1. While the problem of discovering trojans via interpretability tools can be interesting, I am not 100% sure about the motivation behind this work. More specifically what do authors think is the practical use case of such analysis and finding? I think the paper can be motivated better and the introduction can be re-written to reflect a stronger motivation. In general, I believe that writing of the paper can significantly improve as well. 2. The contribution in terms of technical proposal is limited. They propose simple extensions that build on previous work. 3. Section 5.2 could have been explained in more detail. 4. Only a limited number of trojans were tested. Even within the studied trojans the scope was limited to 12 trojans that were implemented. **Minor comments:** Line 68 typo. Line 35 (may comparable -> may be comparable)

Questions

I would like some discussions on motivation in terms of practical use behind this work.

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Yes, authors have addressed limitations behind their work.

Reviewer WymJ2023-08-11

Thanks for your response. I still have some questions.

1. After reading your visualization results about My Q1, If I understand correctly, the higher the correlation score, the higher the success rate of attribution methods in reversing triggers. However, as shown in Figure 2 of one-page PDF, we can observe that kernel-SHAP fails to provide good results. Could you further explain this? If I misunderstood, please correct me. 2. I am sorry that I do not fully understand the Re – Question 4. Could you please further explain it? 3. I recommend the authors discuss Re – Question 5 in the paper if there is any remaining space.

Authorsrebuttal2023-08-11

Thanks

(1) Thank you for pointing this out. We apologize -- there is a mistake in this new plot. The results from ShapleyValueSampling are missing and the three boxes to the left of it are each shifted one to the right (this also causes LIME to have an empty column in the plot). So the best-performing method continues to be Occlusion, but it is mislabeled. We have fixed this, although unfortunately, we cannot share the updated figure. (2) Thanks for the question. Recall that 2 of our 9 methods were based on the visualization of internal neurons. For these, we had to pick a set of internal neurons to visualize, and we did so with a perturbation-based test. We iterated over the network's neurons and perturbed their activations while passing validation data through the network. For each neuron, we multiplied all of its activations by 2 and analyzed how much this affected the output neuron for the trojan's target class in the logit layer. *We then selected the 10 inner neurons that, when their activations were doubled, increased the activations of the neuron for the target class the most on average.* In our answer to question 4, we speculate why visualizing these neurons was not as effective as other methods. We suspect there are two main reasons for this difficulty. The first reason is that networks in general do not detect features (e.g. trojan triggers) via the activation of single neurons. In general, the recognition of neurons is facilitated activation patterns among multiple neurons. The second reason is that even if the recognition of trojan triggers is done by one or a few neurons, visualizing multiple internal neurons may still tend to produce distracting visualizations of irrelevant neurons. We are revisiting our description in the paper to improve clarity. (3) Thank you for the suggestion. We will make corresponding updates to the discussion section of the paper.

Reviewer WymJ2023-08-13

Thanks for your further response.

1. Please make sure that the future version of the content is the correct visualization. 2. About Re – Question 4: Thanks for your answers. I understand how the authors conduct visualizations. However, I think this method could reverse general class-wise semantic patterns for each class rather than specific backdoor features in the target class. Therefore, only based on mentioned methods, I think we may not observe clear backdoor features and they could be easily overridden by class-wise patterns. Perhaps this is also one of the reasons for the failure of visualization. Is it possible that the neurons where backdoor features reside are shared with clean category features? Thanks The reviewer

Authorsrebuttal2023-08-13

Thanks!

1. Absolutely. 2. We agree that the method we use to select neurons to visualize will also select for non-trojan features associated with the target class and that this could be a cofounder. However, we do not believe that this concern should change the general tone of things. (1) Networks typically learn stronger associations between trojan features and the corresponding output than natural features (e.g. [Khaddaj et al., 2023](https://arxiv.org/abs/2307.10163)). And (2) this type of concern is not unique to the inner neuron visualization methods -- any method we test could have ended up visualizing non-trojan features. So we think that it is meaningful that other methods still tend to succeed here while inner neuron visualization fails even. We are updating the text in section 5 to discuss this.

Authorsrebuttal2023-08-11

We fixed a plotting error involving item 2 above

There is a mistake in the new boxplot. The results from ShapleyValueSampling are missing and the three boxes to the left of it are each shifted one to the right (this also causes LIME to have an empty column in the plot). The best-performing method continues to be Occlusion, but it is mislabeled. We have fixed this, although unfortunately, we cannot share the updated figure during the discussion phase.

Reviewer PpyC2023-08-16

Thanks for your helpful responses! Your results here are interesting and would be of interest to the ML community. I will maintain my original rating based on how impactful I believe this paper is.

Authorsrebuttal2023-08-18

Thank you!

Thanks to the reviewer. Please let us know if we could do anything to resolve any questions or concerns between now and the end of the discussion period. Hope you have a nice weekend!

Reviewer EMti2023-08-16

I thank the reviewers for providing nice and detailed explanations and new results. I am willing to increase my score after reading the rebuttal and considering other reviewer comments. As another reviewer was mentioning, writing can be improved significantly to improve clarity including clear motivation behind the work. If authors are going to address this in detail, I am ok with increasing my score. Thanks.

Authorsrebuttal2023-08-17

Thanks for the comment

Thanks! We are glad that some of your reservations have been addressed. We have been making updates for clarity and motivation and will continue our multi-person effort with it. Please let us know if there is anything we can discuss or share (such as an updated version of a particular paragraph) that would help.

Authorsrebuttal2023-08-20

Discussion ends tomorrow

We hope this comment finds you well! We want to follow up to make sure that we can discuss any further questions/comments/concerns. We might be able to run new experiments for the discussion phase but only if we have time before it ends tomorrow. Thanks!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC