Constrained Adaptive Attack: Effective Adversarial Attack Against Deep Neural Networks for Tabular Data

State-of-the-art deep learning models for tabular data have recently achieved acceptable performance to be deployed in industrial settings. However, the robustness of these models remains scarcely explored. Contrary to computer vision, there are no effective attacks to properly evaluate the adversarial robustness of deep tabular models due to intrinsic properties of tabular data, such as categorical features, immutability, and feature relationship constraints. To fill this gap, we first propose CAPGD, a gradient attack that overcomes the failures of existing gradient attacks with adaptive mechanisms. This new attack does not require parameter tuning and further degrades the accuracy, up to 81% points compared to the previous gradient attacks. Second, we design CAA, an efficient evasion attack that combines our CAPGD attack and MOEVA, the best search-based attack. We demonstrate the effectiveness of our attacks on five architectures and four critical use cases. Our empirical study demonstrates that CAA outperforms all existing attacks in 17 over the 20 settings, and leads to a drop in the accuracy by up to 96.1% points and 21.9% points compared to CAPGD and MOEVA respectively while being up to five times faster than MOEVA. Given the effectiveness and efficiency of our new attacks, we argue that they should become the minimal test for any new defense or robust architectures in tabular machine learning.

Paper

Similar papers

Peer review

Reviewer aMTd5/10 · confidence 4/52024-06-22

Summary

This paper aims to study the problem of finding adversarial examples for tabular datasets. Different from attacking image models or text models, attacking tabular models requires finding adversarial examples which are legal, which do not violate the relation between features. Moreover, it also requires to tackle both numerical and categorical features.

Strengths

1. The studied problem is important and under less investigation compared to adversarial attacks in CV and NLP. 2. The paper is well written and easy to follow. 3. The experiment analysis is convincing and sufficient.

Weaknesses

However, I still have concerns on this paper. Specifically, many terms / techniques used in the method are not originally proposed by this paper. For example, 1. In terms of the main objective (as discussed in Section 3.1), one key step to let adversarial examples satisfy the feature constraints (in Eq.1 and Eq.2) is to translate into a differentiable function. However, this is proposed by the previous work Simonetto et al, 2021. 2. In terms of the solving algorithm (as discussed in Section 4.1), the author propose CAPGD, which seems a fine-tuning / lightweight modification of the previous method CPGD. This also limit the technical contribution of this paper. 3. Finally, the author claims trying both CAPGD (proposed in this paper) and MOEVA (previous work) and BF* (previous work) can improve the attack successful rate. Overall, I do not hesitate about the effectiveness of combining these strategies for solving the attack problem. Instead, I am concerned of the originality and the significance of the work. Besides, although the authors discussed the reasons on why differentiable models are considered in this paper, I still believe that undifferentiable models like XGBoost is one important branch of models for solving the tabular data classification tasks. I encourage the authors also investigate possible solutions for undifferentiable models.

Questions

Plz see the weakness.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

I didn't see such a discussion.

Reviewer 5k6D7/10 · confidence 4/52024-07-03

Summary

This paper introduces two novel adversarial attacks, specifically designed to target the evasion of deep neural networks (DNNs) in classification tasks involving tabular data satisfying real-world constraints. The first attack, CAPGD, is a gradient-based method that enhances the constrained PGD (CPGD) attack proposed in [33] by incorporating improvements from the APGD definition [12], such as the addition of momentum and an optimized step size selection. The second attack, CAA, combines the efficient but less effective CAPGD with another search-based attack, MOEVA, which is more effective but less efficient regarding the computational time required to perform the attack. The extensive experimental evaluation on four tabular datasets and five DNNs for tabular data demonstrates that CAPGD significantly improves the success rate of generating successful adversarial examples compared to CPGD. Furthermore, CAA achieves a reasonable trade-off between efficacy and computational efficiency, offering higher attack efficacy than CAPGD and requiring less computational time than MOEVA.

Strengths

I find this paper particularly compelling as it proposes a new evasion attack against deep learning models for tabular data with real constraints that clearly outperforms the other state-of-the-art proposals. The paper demonstrates two key strengths: - Despite the two proposed attacks are improvements and combinations of existing approaches to generate realistic attacks, they are based on reasonable and novel observations. In particular, the design of CAA is motivated by the evidence that CAPGD and MOEVA are complementary. These attacks significantly outperform the state-of-the-art CPGD [33]. - The experimental analysis is comprehensive and convincing. The experiments are performed on 5 different deep learning models employed to classify tabular data on 4 different datasets with variable number of constraints. Both the accuracy and efficiency of the attacks are evaluated. Furthermore, additional results are shown varying the maximum perturbation and the value of important parameters of the attacks are shown. Finally, an ablation study shows the impact of each improvement introduced in CAPGD.

Weaknesses

Even though the proposal is good, I think that some weaknesses about the presentation of the proposal and the analysis of the results need to be addressed in the final version of the paper: **The description of CAPGD should be improved**: the description of CAPGD can be enhanced by better explaining the role of the repair operator. Specifically, the authors mention that constraints of the form $f = \psi$ are enforced at every iteration by the repair operator. However, it is not clear whether applying this operator at every iteration guarantees the generation of an adversarial example that satisfies all dataset constraints and the maximum perturbation constraint by the end of the execution. This is a crucial property, as the attack cannot be tested if it does not satisfy the constraints, potentially wasting the time required to generate the attack. **The depth of the analysis of the results should be improved**: the depth of the analysis of the results can be improved by discussing particular results to understand their causes and how to mitigate these cases. For instance, CAPGD and MOEVA fail to generate adversarial examples on the CTU dataset for 2 out of 5 models (Table 3 and results in the Appendix). This failure may be due to the high number of constraints (360) considered for this dataset, since the two attacks are more effective on the other datasets for which at most 30 constraints are considered. It would be interesting to evaluate the efficacy of the two attacks by varying the number of constraints considered for the CTU dataset, to better highlight the limitations of the proposed attacks. Additionally, CAA generates less effective attacks with higher epsilon values (Figure 5b), and MOEVA generates less effective attacks with a higher number of iterations (Figure 7b). Even though these results are sporadic, they are counterintuitive, as the search space for adversarial examples increases with higher epsilon and more mutation rounds. A detailed discussion of these phenomena would provide valuable insights. **The limitations section is missing**: the paper should include a limitations section summarizing the limitations of the presented attack algorithms and their evaluation. For example, the evaluation is performed only against an $L_2$ attacker, whereas other norm-based or synthetic attackers could have been considered. The attacks may not work well against specific datasets and models when $\epsilon$ is small. While these limitations are reported in the paper, they are scattered throughout. A dedicated section would clarify these points. **Typos and inconsistencies**: there are some typos and inconsistencies in the notation, such as: - $R_\Omega$ in Eq. 3 is not defined when the equation is presented. - $x^{(k+1)}$ -> $x^{(k-1)}$ in Eq. 7. - The classifier is represented by $H$ in Algorithm A.2, but $h$ is used in the main body of the paper. Additionally, $X$ appears twice in the function in line 3 of Algorithm A.2. Correcting these typos and inconsistencies will improve the clarity and readability of the paper. Finally, another weakness is more generic: **The proposed attacks are really specific**: the paper is proposing two effective algorithms to forge realistic evasion attacks against deep learning models used for classifying tabular data, an under-explored setting. I think that the contribution is appreciable. However, the chosen setting may limit the impact of the results, since, at the best of my knowledge, tree-based models continue to be the state-of-the-art for classification tasks on tabular data [a]. [a] Ravid Shwartz-Ziv and Amitai Armon, Tabular Data: Deep Learning is Not All You Need, Information Fusion Volume 81, May 2022, Pages 84-90 **Updates after the authors' response** The authors have provided a satisfying response addressing all the concerns that I have raised as weaknesses. The authors are willing to introduce the findings provided in the response in the next version of their paper, as far as I understood. Moreover, they will also include a dedicated limitations section and correct the typos.

Questions

- Is CAPGD guaranteed to generate an attack that satisfies the dataset and perturbation constraints at the end of execution? - CAPGD and MOEVA may be unable to generate adversarial examples in specific settings (see results on two models on the CTU dataset in Table 3). What is the reason? Could it be due to the high number of constraints of the dataset? How does the success rate of the two attacks vary considering different numbers and types of constraints (linear and nonlinear)? - Why may CAA and MOEVA show a smaller success rate when considering higher epsilon values and a higher number of rounds (see Figures 5b and 7b)?

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

The authors should sum up the limitations of their proposals in a specific section in the paper. The negative societal impact is discussed by the authors who claim that their work may give birth to new and stronger defenses. Finally, the experimental and implementation details have been extensively documented.

Reviewer 5k6D2024-08-08

Compliments to the authors for their high-quality response.

I sincerely thank the authors for their wide and deep response. Their points sound reasonable and clarify the open points that I raised in the review. I hope that the authors will discuss their new results in the next version of their paper. I have only an observation about the results in Table 3 of the attached PDF. It seems to me that attacking STG is difficult independently of the considered constraints since its robust accuracy is always 95.3%. Understanding if the model is really robust or if a better attack is needed could be an interesting future work. I will modify my review to acknowledge your response and improve my score, given the depth of your analysis and dedication you have shown in your response.

Authorsrebuttal2024-08-13

We thank you for your praise of our rebuttal and we are happy to see you increase your score following our answers. To answer your observation on Table 3 on the author's response PDF. We confirm that STG architecture is the hardest to attack both with gradient and search attacks. We show consistent robustness of STG in Table 3 of our main submission across all the datasets. STG [A] is a novel architecture that implements an embedded nonlinear feature selection method by introducing the stochastic gates to the input layer (the feature space) of a neural network. The randomness introduced on the fly to select the features for training and inference significantly hinders evasion attacks, both gradient and search attacks. There is a similar phenomenon in the benchmark by Croce et al. for computer vision (Robustbench). In their evaluation of AutoAttack, they clearly "rule out classifiers which have (1) zero gradients with respect to the input, (2) randomized classifiers, and (3) classifiers that use an optimization loop at inference time" [B]. Contrary to their benchmark, we decided to cover one architecture per family of mechanisms from the Tabular ML literature, including stochastic mechanisms (STG belongs to categories 2 and 3 discarded in Robustbench). In Table 3 of our main paper, while we demonstrate that STG is the most robust, our study uncovers some cases where STG robust accuracy could still be significantly decreased with CAA (LCLD, URL), while other datasets will be challenging targets for future research (CTU). We appreciate your feedback on this pattern and the discussion it raises. In the appendix, we described the mechanisms of each architecture but did not connect the mechanisms with the observed robustness. Following your insight, we will also update this section to explain how each mechanism could impact the robustness (a complete verification of each mechanism impact would be a natural follow-up for dedicated papers). Thank you again for this discussion and the points you raised. They have significantly improved our final version. [A] Yamada, Yutaro, et al. "Feature selection using stochastic gates." ICML, 2020. [B] Croce et al. "Robustbench: a standardized adversarial robustness benchmark."(NeurIPS, 2021).

Reviewer 5k6D2024-08-13

Official Comment by Reviewer 5k6D

I appreciate your investigation, which provides an intuition about the robustness of STG. Adding these observations to the paper will certainly strengthen your contribution. Thanks again; I will continue the discussion with the reviewers.

Reviewer JEHw6/10 · confidence 4/52024-07-09

Summary

The paper proposes two adversarial attack methods targeting deep learning models for tabular data. The two methods are: CAPGD (Constrained Adaptive Projected Gradient Descent) and CAA (Constrained Adaptive Attack). CAPGD is a modification based on constrained PGD with step size adjusting, repair operator, additional random initialization, and momentum. CAA is a combination of CAPGD and MOEVA (Multi-Objective Evolutionary Adversarial Attack), which is a search-based attack method. The two methods are combined by iteratively applying the two with CAPGD first. The authors demonstrate the effectiveness of these attacks across five architectures and four datasets.

Strengths

Strengths: 1. Clarity: The motivation and rationale behind the proposed algorithms are clearly presented in a logical order. 2. The paper provides an extensive empirical evaluation of the proposed attacks across multiple datasets and architectures, showcasing their superiority in terms of effectiveness and computational efficiency.

Weaknesses

Weakness: 1. Lack of novelty and contribution: CAPGD is modified base don CPGD with a series of commonly used optimization techniques. CAA is a combination of CAPGD and MOEVA, which was proposed by a previous work. 2. The implementation of CAA, involving the combination of CAPGD and MOEVA, might be complex and resource-intensive, which could limit its practical applicability in some settings. 3. Lack of evaluation against defense mechanisms: Section 5.4 briefly discussed potential defense mechanisms, Madry’s adversarial training, which was proposed six years ago. There have been many works proposed since then. It is not fair to say that it is the only reliable defense against evasion attacks. Minor: • On page 1, line 30, “This raises anew the need to study …”. Maybe it should be “This raises a new need to study …”?

Questions

NA, see above.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

No discussion found for limitation.

Reviewer XFcC7/10 · confidence 3/52024-07-11

Summary

This paper considers the evaluation of the robustness of deep learning models applied to tabular data. The authors introduce two novel adversarial attack methods: Constrained Adaptive Projected Gradient Descent (CAPGD) and Constrained Adaptive Attack (CAA). These methods are designed to exploit the vulnerabilities of tabular data models, which often include categorical features, immutability constraints, and feature relationship constraints that are not typically considered in attacks designed for computer vision (CV) or natural language processing (NLP). CAPGD improves on existing gradient-based attacks by incorporating adaptive mechanisms and eliminating the need for parameter tuning, significantly enhancing the success rate and efficiency of generating valid adversarial examples. CAA further combines CAPGD with the Multi-Objective Evolutionary Adversarial Attack (MOEVA), optimizing both effectiveness and computational cost. The paper demonstrates the superior performance of CAA across various datasets and models, suggesting it should become a standard test for evaluating the robustness of tabular models.

Strengths

- The paper is well-motivated and logically structured, contributing significantly to advancing adversarial machine learning, particularly in the domain of tabular data. - The proposed CAA method is SOTA on effectiveness and efficiency, making it a good benchmark for testing the robustness of tabular data models, similar to the role of AutoPGD/AutoAttack in computer vision tasks. - The paper includes extensive and systematic experiments, providing solid empirical evidence for the effectiveness of the proposed attacks.

Weaknesses

- The paper does not provide a clear explanation of the penalty function in Equation (4). A more detailed clarification is needed to understand how this function is formulated and applied. - The description of constraints in Equations (1) and (2) is somewhat abstract and difficult to interpret. The authors should refine the descriptions of these constraints to improve clarity.

Questions

- Are there any other constraints in tabular data besides categorical features, immutability, and feature relationship constraints? - Why are gradient attacks ineffective on the CTU dataset? - Why is feature engineering necessary for the datasets? Have all four datasets considered in the paper undergone feature engineering? It seems that the WiDS dataset has not undergone feature engineering. Is that correct?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

None

Reviewer aMTd2024-08-11

We thank the reviewer's response. I agree with the reviewer that this method is potential to be served as a good baseline strong attack in the related literature, thanks to the authors' effort on trying different strategies for performance improvement. It is also insightful to see the transferability of attacking between different types of models. Thus, I increase the rating to 5.

Authorsrebuttal2024-08-13

We thank you for your response and positive feedback on our extended analysis and new results. We appreciate that you have increased your score to 5 and are happy to provide any additional insight or answer any interrogation you may have to fully satisfy your requirements for a clear acceptance of our work.

Authorsrebuttal2024-08-14

Summary of our improvements

Dear reviewer, we thank you again for your constructive feedbacks, and your support of our work. We would like to highlight that we have addressed during this rebuttal all your concerns and interrogations and we have provided additional insights and results to support our claims. In particular, we have adressed your interrogations as follows: * Q1: Are there any other constraints in tabular data besides categorical features, immutability, and feature relationship constraints? => A1: To the best of our knowledge, our work covers all the constraints related to tabular data. * Q2: Why are gradient attacks ineffective on the CTU dataset? => A2: We provided a detailed analysis in the common answer with an analysis of the constraints of CTU. CTU is robust because of the number of constraints and the number of features involved in each constraint. * Q3: Why is feature engineering necessary for the datasets? => A3: We did not run any additional feature engineering on the datasets. Each dataset was designed by a different source and may have required dedicated feature engineering. We were not involved in this step. If you find our answers and discussion satisfactory, we would greatly appreciate it if you could increase your score accordingly. If there are any remaining issues or questions, we would be more than happy to address them before the discussion period ends. Thank you again for your insights and the discussion points you raised.

Reviewer XFcC2024-08-14

I thank the authors for the detailed response. My concerns have been well resolved. In addition, I appreciate the additional experimental results reported in the rebuttal, which demonstrate the practicality of the proposed attacks. Therefore, I would like to increase my score from 6 to 7 accordingly.

Authorsrebuttal2024-08-14

Summary of our improvements

Dear reviewer, we thank you again for your constructive feedbacks, and your suggestions to improve our work. We would like to highlight that we have addressed during this rebuttal all your concerns and interrogations and we have provided additional insights and results to support our claims. In particular, we have adressed the weaknesses you raised as follows: * W1: Lack of novelty and contribution. => A1: We have elabotated on the complex process of desiging our new CAPGD attack, that required analyzing multiple improvement mechanisms, their interactions, thus leading to only including in CAPGD the most relevant mechanisms. Our CAPGD design is the best that could be achieved for gradient attacks in tabular ML. We also detailed the inception of our new meta-attack, CAA and explained why its design is not straightforward and required the analysis, selection and evaluation of 10 existing attacks. * W2: The implementation of CAA might be complex and resource-intensive => A2: We elaborated on our previous analysis on the computation cost and execution time of CAA compared to MOEVA (Table 3 of our manuscript) and explained that CAA is marginally more ressource-intensive than MOEVA in one over 4 datasets and significantly more efficient than MOEVA in the 3 remaining datasets. * W3: Lack of evaluation against defense mechanisms => A3: We have implemented 5 new defenses using extensive data augmentation and adversarial training. Our new defenses leverage some of the best and most recent generative models of tabular data and required the training of complex generative models to generate 100 times more synthetic examples and achieve the best adversarial training defenses. Our new extensive experiments show that these new defenses can significantly improve the robustness of the models to CAA, but that our new CAA attack remains effective for URL and LCLD datasets across all architectures, and for WIDS on TabTransformer and STG architectures. If you find our answers and discussion satisfactory, we would greatly appreciate it if you could increase your score accordingly. If there are any remaining issues or questions, we would be more than happy to address them before the discussion period ends. Thank you again for your insights and the discussion points you raised.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC