Regularization properties of adversarially-trained linear regression

State-of-the-art machine learning models can be vulnerable to very small input perturbations that are adversarially constructed. Adversarial training is an effective approach to defend against it. Formulated as a min-max problem, it searches for the best solution when the training data were corrupted by the worst-case attacks. Linear models are among the simple models where vulnerabilities can be observed and are the focus of our study. In this case, adversarial training leads to a convex optimization problem which can be formulated as the minimization of a finite sum. We provide a comparative analysis between the solution of adversarial training in linear regression and other regularization methods. Our main findings are that: (A) Adversarial training yields the minimum-norm interpolating solution in the overparameterized regime (more parameters than data), as long as the maximum disturbance radius is smaller than a threshold. And, conversely, the minimum-norm interpolator is the solution to adversarial training with a given radius. (B) Adversarial training can be equivalent to parameter shrinking methods (ridge regression and Lasso). This happens in the underparametrized region, for an appropriate choice of adversarial radius and zero-mean symmetrically distributed covariates. (C) For $\ell_\infty$-adversarial training -- as in square-root Lasso -- the choice of adversarial radius for optimal bounds does not depend on the additive noise variance. We confirm our theoretical findings with numerical examples.

Paper

Similar papers

Peer review

Reviewer uRwm8/10 · confidence 4/52023-06-13

Summary

This paper provides an in-depth analysis of adversarial training with linear models and its relationship to regularized regression methods under the overparametrized regime. Depending on the value of the perturbation radius, it is revealed that there exist three modes. When the radius is small, solutions to adversarial training behave as minimum-norm interpolators (Theorem 1). When the radius is medium, the solutions behave as solutions to the parameter shrinkage regression (Proposition 4). When the radius is large, the zero solution is necessary and sufficient (Proposition 3). In addition to these theoretical results, the authors observe the mode change experimentally and discuss how adversarial training is advantageous over parameter shrinkage regression.

Strengths

- A modern extension of theory on robust optimization and regularization: The relationship between robust optimization (somewhat encompassing adversarial training in this work) and regularization has been known in the literature, including Xu et al. (2009). This work contributes to studying what happens when it comes to overparametrization and nicely characterizes the relationship between the perturbation radius and the corresponding modes (as I summarized above). - Demonstration of the benefit of overparametrization: In the numerical simulation of Figure 2, the authors demonstrate that the robustness radius increases as the model becomes more overparametrized, namely, $p/n$ increases. This clearly indicates the benefits of overparametrization (though the analysis hinges on norm matching, as mentioned in Remark 2). - Clarity: Despite the thorough theory, the paper is written clearly and easy to follow. Xu et al. (2009). "Robustness and Regularization of Support Vector Machines." (JMLR)

Weaknesses

One of the main weaknesses would be the restriction to linear models, which is crucial for the current analysis yet needed for further understanding adversarial training. You may refer to Xu et al. (2009) when you show Theorem 4. Indeed, the equation right after l.322 can be regarded as a generalization of Theorem 3 in Xu et al. (2009) because $\\ell(y(\\boldsymbol{x}^\\top\\boldsymbol{\\beta}) - \\delta\\|\\beta\_\*\\|) \\le \\ell(y(\\boldsymbol{x}^\\top\\boldsymbol{\\beta})) + \\delta\\|\\boldsymbol{\\beta}\\|\_\*$ when $\\ell$ is the hinge loss. Below, I have other minor comments. - In Figure 1, can you specify what $\\lambda$ and $\\delta$ are used for each line? - In the proof of Theorem 1, you may need $-$ (negative) sign in front of either $\\epsilon\_i\boldsymbol{x}\_i$ in Eq. (6) or $n\\delta\boldsymbol{\\alpha}$ in l.132. Otherwise, "the subderivative contains zero" (l.132) does not seem to be correct. - In l.150, the reason of $\\delta\_{\\text{test}} \\propto \mathbb{E}[\\|\\boldsymbol{x}\\|]$ is unclear to me. Can you elaborate on it? - In Figure 4, can you specify what $n$ and $p$ are used? - In Eq. (8), do you miss the exponent $2$ for the norm? - In Eq. (9), it might be better to change the notation $\\epsilon$ for the noise because $\\epsilon$ has already been used in the proof of Theorem 1. - In the equations after l.319 and l.322, should we need $+ \\Delta x$ on the left-hand sides? - In the appendix, what is referred to as Theorem 3 seems to be Proposition 3.

Questions

See the weaknesses.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

4 excellent

Contribution

3 good

Limitations

Obviously, the analysis is entirely limited to the linear model case. Nonetheless, the analysis provides a fair amount of insights to readers, so I don't think this is a big limitation.

Reviewer Zg7q6/10 · confidence 4/52023-06-16

Summary

This paper studies the connection between adversarial training and regularization methods in linear regression problem. Simulation studies are provided to justify the correctness of their theoretical observations.

Strengths

The authors conducts a comprehensive study on the relationship between adversarial training and regularization methods in linear regression setup. The writing is clear and easy to understand.

Weaknesses

My major concern towards this paper is the limit of its contribution. While the analysis is comprehensive, it is only restricted to linear models. Considering that the adversarial training is more commonly used in neural networks rather than linear models in reality, the contribution is limited. The authors are encouraged to add more discussions on neural networks. In addition, the following paper considers the connection between regularization and adversarial robustness: Jakubovitz, Daniel, and Raja Giryes. "Improving dnn robustness to adversarial attacks using jacobian regularization." Proceedings of the European Conference on Computer Vision (ECCV). 2018. Please cite this paper and compare it to the submission from intuition aspect.

Questions

Is it possible to extend the analysis to two-layer neural networks?

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

NA

Reviewer 6WVQ6/10 · confidence 3/52023-07-03

Summary

The paper studies adversarial training (AT) for linear regression for which the inner maximization problems has a closed form solution. They then attempt at relating the solutions to solutions of other optimization problems: - They show that the minimum norm interpolator also minimizes the adversarial loss (iff the adversarial perturbation is sufficiently small) - They show that adversarial training (under certain conditions) minimizes something closely related to the LASSO and ridge regression objective for $\ell_\infty$ and $\ell_2$ attacks respectively. - They show that similarly to square-root LASSO, adversarial training does not need knowledge of the variance and they argue that this makes adversarial training a viable alternative.

Strengths

- It seems interesting to attempt connecting AT to sparse solutions - The initial setup and the statements of the theorems are presented in a clean way - Existing literature is well-covered

Weaknesses

- My main concern is that the theoretical claims are rather weak: - Concerning Thm. 1, l. 122 "minimum-norm interpolators as the outcome of adversarial training" seems a bit of a stretch, since it is not *consistently* the outcome of adversarial training (we might be able to find a minimizer of $R^{adv}$ that is *not* a min norm interpolator). AT would imply minimum-norm interpolator if the minimizer of $R^{adv}$ was unique, but this cannot be the case since LASSO is not unique in general. - Prop. 2 is concerning minimum norm interpolator (so not necessarily obtainable with AT!). What makes this statement interesting for adversarial training if we need to obtain the solution through other means? - Prop. 4 seems to not directly relate AT to LASSO/ridge regression. Whats is the conclusion of Prop. 4? - Thm. 2 exists to show that AT can replace sqrt-root LASSO. You are comparison with Lasso though – doesn't the bound have a bias in comparison with sqrt-root LASSO (eq. 10 of [29]). The main feature of AT seems to be the claim that $\delta^*$ is invariant to rescaling of $\varepsilon$. Can you explicitly make $\delta^*$ in Thm. 2 independent of $\varepsilon$? (currently this is not the case in theorem statement) Comments: - Prop. 5 maybe pick a different variable than $p$ (already used for dimensionality) - l. 144 should have been $\delta$ instead of $\delta_{train}$? - Maybe write "a solution" in l. 144 instead of "the". - l. 179: Please describe the dataset in the appendix or provide a more direct pointer to [18].

Questions

- Thm. 1: $\bar \delta$ depends on the $\ell_\infty$-norm regardless of the choice of norm in the adversarial training? This seems potentially loose – could you comment on it? - Prop. 2: Do you still rely on full row rank in Prop. 2? - l. 158-159: Isn't the claim in [17] about $\ell_2$ minimum norm while your Prop. 7 is a claim about choice of norm in the adversarial training? - Figure 3: Could you label the plot to explain the colors? I don't understan how to interpret the plot. - Figure 4 / l. 179: what is "regularization paths"? - What assumption breaks in Prop. 4 since it is no longer able to predict similarly after $\delta$ is made sufficiently small (as demonstrated in Fig. 4)?

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

N/A

Reviewer CYXN6/10 · confidence 4/52023-07-04

Summary

This paper investigates adversarial training of linear regression. The authors compared the solution of adversarial training and other regularization frameworks (minimum-norm interpolating, ridge regression, Lasso and square-root Lasso), and established close relations between adversarial training and other methods under certain conditions depending on the disturbance radius and over/under-parameterization. The authors also consider extending the result to more general loss function for linear model.

Strengths

1. The paper provides valuable insights on the relation between adversarial training and other regularization frameworks for linear regression, which contributes to the area of robust learning. The analysis is sound. 2. The paper provides good background knowledge and details in their work. 3. The paper is well-organized and easy to follow overall.

Weaknesses

1. In the abstract, the authors claim that adversarial training can be equivalent to parameter shrinkage methods (like ridge regression and Lasso). However, from Proposition 4, it seems the two frameworks are not equivalent, since the regularization term in the equation of Proposition 4 equals $\delta^2\left\| \beta\right\|^2 + c\delta\left\| \beta\right\|$ for some constant $c$. I am curious about how the quadratic term can affect the solution, or how close the adversarial training solution is from the parameter shrinkage solution. 2. In the numerical experiment, the authors have not mentioned how the adversarial training is carried out in these datasets. From the code in the supplementary materials, it seems the adversarial samples are generated by the PGD attack. Please consider including more details in the paper. Also, does PGD generate sufficiently strong attacks for linear regression?

Questions

Here are some additional questions/comments: 1. $\sigma_1$ and $\sigma_n$ in line 126 are undefined. 2. The authors claim in line 151 and Remark 2 that the model becomes robust as feature dimension $p$ grows, which seems not precise to me. The authors suggest that the threshold $\bar{\delta}$ increases faster, but this only guarantees that the optimal solution of adversarial training and minimum-norm interpolator agree, which does not necessarily mean more robustness. Does the risk $\mathcal{R}^{\text{adv}}$ decrease as feature dimension $p$ grows? 3. The paper investigates the situation where the sample features $x_i$’s are disturbed in linear regression. In applications, it is also very common that the target $y_i$’s are disturbed.

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

The authors have adequately addressed the limitations.

Reviewer uRwm2023-08-15

Response

Thank you for the response. The authors' elaboration mostly cleared up my questions. On my minor comment > In the equations after l.319 and l.322, should we need $+\\Delta x$ on the left-hand sides? I just wanted to draw attention to possible typos in the left-hand sides; for example, $\\max_{\\|\\Delta x\\| \\le \\delta}\\ell(|x^\\top \\beta-y|)$ (right after l.319 in the original version) seems to be $\\max_{\\|\\Delta x\\| \\le \\delta}\\ell(|(x+\\Delta x)^\\top \\beta-y|)$, otherwise there is no $\\Delta x$ to be maximized, which does not align with the left-hand side of eq. (11). If I still misunderstand anything, feel free to point out.

Reviewer 6WVQ2023-08-17

Thank you for the elaborate response which clarifies a lot. I'm mostly happy with the response and have raised my score, even though I still remain convinced that the theoretical statements regarding AT are rather weak. I would just ask you to include the following: - Thm 1 & Prop 2: Since uniqueness is not established in Thm 1, we cannot use the statement to draw conclusions about AT from characterization of minimum-norm interpolators (Prop 2). Because of that, apart from toning down the claim regarding Thm 1, I would maybe also explicitly clarify that Prop 2 is not characterizing AT (since AT is otherwise the focus of the paper). - Prop 4: The last equation you obtain in your rebuttal after rewriting the proposition contains $\|\beta\|_1$. I guess this should have been $\|\beta\|_\infty$. Including the form that you provided in your rebuttal and explicitly commenting on $\lambda'$ would make the connection more precise.

Authorsrebuttal2023-08-20

Thank you very much for taking the rebuttal into consideration and raising your score. Also, thank you for the two suggestions: we will clarify Proposition 2 and add the connection in Prop 4 (with the ideas we wrote in the rebuttal).

Reviewer CYXN2023-08-18

Thanks much to the authors for the detailed reply. Most of my questions are addressed. For the relation with Lasso, it seems the value of $\lambda^\prime$ will additionally depend on $\|y\|_1$, which might require more effort to tune compared with the standard Lasso framework. Also, other reviewers raise some concerns that I did not think of before. Overall, I tend to keep my previous rating.

Authorsrebuttal2023-08-20

Again thank you very much for all the comments and the careful review. A small clarification: in the case of adversarial training you tune $\delta$ which might be easier to adjust than the standard Lasso parameter since it does not depend on the noise amplitude (see Section 6). The value $\lambda'$ would not be accessed by the user of the method and would not be adjusted directly.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC