acceptodds
Withdrawn from ICLR 2027

Why Robust Classifiers Can Generate: Hidden Denoising structure

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Adversarially robust neural networks, while designed for classification, exhibit surprising generative capabilities when appropriately probed. We provide a theoretical framework explaining this phenomenon by connecting adversarial robustness to implicit denoising structure. Building on established results that robust training drives Jacobians toward low-rank solutions, we demonstrate that the Gram operator functions as an implicit denoiser, selectively preserving dominant directions while suppressing noise in orthogonal directions. We provide a mechanistic explanation for emergent generative capabilities in robust classifiers, validated through a diagnostic probing procedure that makes the denoising operator empirically observable. Our results establish a connection between discriminative robustness training and generative inference, showing that robust classifiers encode statistical priors that enable structured pattern generation without explicit generative objectives.

Primary area:
generative models
Keywords:
Image generation, Adversarial robustness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_HKTEOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates the capabilities of a robust classifier to generate images by inverting the classifier, i.e., by examining the gradient of the loss with respect to the input. The paper discusses the Jacobian JJ instead of the gradient when differentiating the network across all classes. The paper claims that the matrix J⊤JJ^\top J is an implicit denoiser that suppresses noise in orthogonal directions. The article reports a proposition and a theorem that look almost obvious to a researcher who is an expert in this field. The theorem is of interest, but it is also almost a consequence of the fact that J⊤JJ^\top J is of low rank. From them, at least in my opinion, it is hard to derive new information. Perhaps the most interesting part is Eqs 6-9, where a proxy is used to compute J⊤JJ^\top J implicitly, but the approach still receives the gradient from the network, as in all previous literature. Also, that part works by linearizing the network. In terms of experimental results, I do not see anything new there: the experiments confirm the hypothesis of the paper, but do not add much to the literature. Their report numbers related to the implicit denoiser hypothesis and the spectral statistics, and how these models can generate images from random noise (known). The paper kind of tries to do it at multiple scale (CNN, Transofmers), different norms, different layers, but it does not offer a full ablation on these and a full study.

Critical strengths
  • The only interesting part to me is Eq. 6-9, where the J⊤JJ^\top J is linked to take the gradient from the network providing an input image xx and its perturbed versions x+ϵx+\epsilon using stop gradient. That part (modulo the linearization) is cool, and I learned something.

  • Theorem 1 is minimally interesting, but it just derives from applying the low-rank structure of J⊤JJ^\top J to a distribution as input ϵ ∼N(0,σ2IP). I would not count this as strong novelty.

Critical weaknesses
  • The community has published a lot of papers on the mystery of robust classifiers acting as generators. The paper reports most of them, with some missing. see below [A, B, C, D]. The main problem is that by reading the paper, I was not able to get new knowledge on the topic. I believe this topic is highly saturated, so gaining new knowledge or achieving a breakthrough is very difficult. The reason is that everyone works either at the output level (logits) or the input level (Jacobian input, etc.), but no one dares to understand what happens to the nonlinear part. Bottom line: contributions are a bit too weak; more or less, it sounds the same, but with different words.

  • On the claim that J⊤JJ^\top J is an implicit denoiser, this is true even for a matrix U⊤UU^\top U where UU are the principal components of a dataset with decay spectrum (natural images). If you pass data into U⊤UU^\top U, this will reconstruct images by just passing the low-frequency part. So if J⊤JJ^\top J is low-rank (prior work), I am not surprised by the finding that it acts as a denoiser.

  • Some claims are not correct. for example,

sentences like this

structure without explicit denoising training. [about adv. training]

They are not entirely true. Diffusion Models (DM) inject noise and then propagate noise in the loss (equivariance) Robust Classifier also injects noise (which is a function of the network) and then applies invariance (due to classification needs) So they are doing some denoising training; the problem is that, given that this was approached by a classification/discriminative model, people did not realize it. Every time you perturb the data somehow, you go back and look at the input density (i.e., you do generative modeling).

Theorem 1 does not add much; the PSD matrix J.T@J is already low-rank and thus acts as a denoiser. The proof and Theorem 1 only make this considering the input to be Gaussian noise. does not add up much. In other words, Theorem 1 puts the projection/unprojection of a generic point input xx for a distribution. It is known that components ortogonal to the span of J do not survive after J.T@J.

Given this spectral structure, J⊤J acts as a denoiser.

Regarding your claim. The fact that J⊤J acts as a denoiser is known if as low rank-structure and decay spectra.

The probing objective reduces the linear network to a linear operator, which is oversimplistic. i.e., gradient over the input is the Gram operator only if the network is linearized! This is an important distinction.

Eq. (6) is nothing new; it is essentially the same as optimizing the input to increase the logit of the desired class k. In the case of increasing the logit of the desired class k, the reference embedding r(x)r(x) will be the k-vector in the classification layer. In equation (6), it is replaced with the embedding of natural data. So, even in this part, I do not see new things.


  • On the paper quality part:
    • title should contain a ? (question mark)
    • A lot of figures are there, but I am already on page 6, NO figure is cited in the text! The first figure is cited on page 7. Figure 2 is not cited.

The article is missing several related works:

  • [A, B] similar generation approach reported in this paper.
  • Connection to EBM [C]
  • [D] with other strong priors, we can generate from discriminative models even with non-robust encoders.

I strongly believe that if you send this paper to ICLR, you cannot miss papers like [A] and [B]


[A ] Towards understanding the generative capability of adversarially robust classifiers. ICCV 2021

[B] Shedding More Light on Robust Classifiers Under the Lens of Energy-Based Models. European Conference on Computer Vision (ECCV), 2024

[C] Understanding Adversarial Training with Energy-based Models. https://arxiv.org/pdf/2505.22486

[D] Implicit Inversion turns CLIP into a Decoder. ICLR 2026.

Questions for authors

Theorem 1 takes spectral concentration as a premise. We verify empirically that it holds in adversarially trained networks (Section 5);

Did prior work already prove this empirically?

matrix multiplication requires O(P^2*K)

Is this really a bottleneck? Can you approximate somehow Jk.T@Jk for one class, this is O(P^2) Then do this in parallel, and then sum across all classes? Maybe in CIFAR10 this could work.

I do not see why I should buy that L2 is better than L_inf just by looking at one sample in Fig. 3B

Reviewer_TqhROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper provides the analysis that why implicit denoising structures can help boost the adversarial robustness in deep learning. They that the Gram operator J⊤J\mathbf{J}^{\top} \mathbf{J} functions is an implicit denoiser. It can selectively preserve dominant directions while suppressing noise in orthogonal directions. The paper further proposes a mechanistic explanation for emergent generative capabilities in robust classifiers, validated through some empirical studies. This connects adversarial robustness and generative inference.

Critical strengths
  1. Theorem 1 is relatively rigorous in mathematical description.
  2. The design of Section " EMPIRICAL VALIDATION VIA J⊤J\mathbf{J}^{\top} \mathbf{J} PROBING" is very interesting. It satisfies the standard of ICLR to design some experiments for insights.
Critical weaknesses
  1. This paper uses "spectral amplitude reduction" to explain generation, but the actual algorithm normalizes this factor. The reduction of the overall spectral amplitude cannot be solely explained by the almost invariant generation process of overall scaling. This does not deny that robust models do generate more structured images, but it means that the paper needs to prove whether it is the normalized spectral shape, semantic orientation of the main subspace, or other nonlinear factors that really work, rather than relying mainly on "small gain, small trace, small tail" to argue.

  2. The equation of Theorem 1 basically holds, but it does not prove the usual meaning of "denoising". Reducing noise energy does not mean restoring useful signals; Low rank itself does not have semantic guarantee; The low rank caused by K≪PK \ll P also holds true for standard classifiers.

  3. From "classification robustness" to raw logit Gram structure, there are still unresolved unidentifiable issues.

  4. The experiment has not yet connected "structured visualization" with "explanation generation mechanism". The robust classification experiment was only conducted in MNIST and lacks persuasiveness. There should be more diverse benchmark data available. Especially the spectral analysis of Imagenet was conducted, but there is no explanation of its noise resistance performance on this dataset.

  5. Why this paper only implements ℓ2\ell_2 norm AT?

  6. What is difference between this paper and previous work?

Santurkar, Shibani, Andrew Ilyas, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. "Image synthesis with a single (robust) classifier." Advances in neural information processing systems 32 (2019).

Engstrom, Logan, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. "Adversarial robustness as a prior for learned representations." arXiv preprint arXiv:1906.00945 (2019).

Zhu, Yao, Jiacheng Ma, Jiacheng Sun, Zewei Chen, Rongxin Jiang, Yaowu Chen, and Zhenguo Li. "Towards understanding the generative capability of adversarially robust classifiers." In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 7708-7717. IEEE, 2021.

Questions for authors

Same as weakness.

Reviewer_PSnXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to explain why adversarially robust classifiers can generate images. The authors argue that J^T J acts as an implicit denoiser when its spectrum is concentrated, and support this with spectral measurements and a probing procedure that generates images from noise. The question is important, but the theory does not distinguish robust from standard models, and the experiments mostly restate known results.

Critical strengths

The question is important and the direction is promising. Why adversarial training leads to perceptually aligned gradients and generative behavior is an open question. Connecting this to denoising and diffusion is an interesting direction, and while reading the paper I genuinely hoped to find an answer. Unfortunately, as discussed below, I do not think the paper provides one.

Critical weaknesses

W1. The theory adds no information beyond basic linear algebra, and the paper should not claim it as a theoretical contribution. Theorem 1 says only this: if the tail eigenvalues of J^T J are small, noise along those directions is suppressed. This is true of any matrix, and since rank(J^T J) ≤ K, it holds for every classifier. The theorem cannot say that adversarial training produces a better denoiser than standard training. It cannot say why the surviving directions look like images either. Both questions are pushed outside the theory, as a premise and as an empirical check. The theorem also depends only on the shape of the spectrum, not its scale. So the large magnitude gaps that the experiments keep emphasizing, such as trace, energy ratio, and tail eigenvalues, mean nothing for the theorem. And the shape is similar for robust and standard models (soft-rank in Table 2).

The paper itself seems to recognize these limits, as strong claims are often followed by a step back. The abstract states that J^T J “functions as an implicit denoiser”, while Sec. 3 takes spectral concentration “as a premise”. The conclusion claims to reveal “the mathematical structure underlying” generation, while the limitations state that causation is not established. To me, this reads more as framing than as theory. I think the claims of a theoretical contribution in the abstract, introduction, and conclusion should be removed, and the analysis positioned as an interpretation that offers some insight.

W2. Without the theory, there is little left to evaluate. The paper asks why adversarial training leads to perceptually aligned gradients and generation. Once Theorem 1 is read as intuition, the experiments only show that robust classifiers have perceptually aligned gradients and can generate images, both of which are already well known. The value of the paper was supposed to lie in explaining why. Since it does not, I am not sure what contribution is left to evaluate.

Questions for authors

Suggestion. The key question seems to be why adversarial training, unlike natural training, makes the class gradients meaningful. Theory on how adversarial training changes the learned features, such as Feature Purification (Allen-Zhu & Li, 2021), could be a good starting point, and I would be glad to see this question answered.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.