acceptodds
Withdrawn from ICLR 2027

PMF-Purify: Unifying Accuracy and Efficiency in Single-Step Pixel-Space Adversarial Purification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep neural networks remain vulnerable to imperceptible adversarial perturbations. Adversarial purification provides a robust, plug-and-play defense by restoring perturbed inputs prior to classification without modifying the target model. While diffusion models offer powerful natural-image priors for this task, their iterative reverse process severely bottlenecks inference speed. Current acceleration methods largely attempt to compress these diffusion trajectories via distillation or guided few-step denoising. We argue for a paradigm shift: because adversarial examples inherently retain the vast majority of their original visual semantics, progressively regenerating them through a multi-step trajectory is fundamentally unnecessary. In this work, we introduce PMF-Purify, a single-step, latent-free adversarial defense operating in pixel space. Building on Pixel MeanFlow (pMF), PMF-Purify bypasses iterative sampling by formulating purification as a direct endpoint prediction. By utilizing a parameter-efficient LoRA adaptation—keeping the base generative prior and downstream classifier entirely frozen—it maps a perturbed intermediate state directly to the natural-image manifold in exactly one network evaluation (NFE = 1). Under the rigorous ImageNet-1K/ResNet-152 evaluation protocol, PMF-Purify achieves a state-of-the-art average robust accuracy of 78.6% across 11 diverse attacks. Strikingly, it accomplishes this with an inference latency of merely 3.72 ms per image, successfully unifying top-tier adversarial robustness with single-digit millisecond efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
adversarial examples, adversarial purification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_56UcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This job is mainly dedicated to solving the problem of slow inference speed in adversarial purification processes based on generative models. The author provided experimental results on ImageNet-1K and CIFAR-10, and theoretical analysis validated the relevant claims.

Critical strengths
  1. The main experiments in this article are well conducted, verifying the robustness of ResNet-152 and WideResNet-50-2.
  2. The effectiveness of the method was further qualitatively analyzed through interpretability experiments.
Critical weaknesses
  1. This article only conducted experiments on CNN, not on ViT.
  2. This article only includes BPDA-PGD experiments, not BPDA-EOT experiments.
  3. The attack cost and number of steps in this article are too small, especially the number of attack steps.
  4. The AA results should be compared on a standard benchmark.
  5. No comparison was made with AT.
  6. The experiment can be implemented on both ℓ∞\ell_\infty and ℓ2\ell_2-norm attack settings.
  7. Table 6 shows that the purification strength hyperparameter ss is highly sensitive, with inherent trade-offs and no adaptive mechanism. In recent years, there have been multiple purification works. This paper proposes a plugin, but it should demonstrate its universality rather than simply adjusting parameters to obtain results.
Questions for authors
  1. Could the experimetns on ViT be supplemented on the maanuscript?
  2. How could you explain the significant attenuation of robustness under strong adaptive white box attacks?
  3. Why does this article avoid the gradient masking problem and dare not directly compare with AT?
  4. What does the experimental result in Figure 4 want to prove?
  5. Inference heavily relies on Gaussian noise sampling, introducing random factors. Why did you only conduct one experiment instead of multiple sub-experiments?
Reviewer_Ver5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to speed up diffusion-based adversarial purification methods. The proposed PMF-Purify leverages a pretrained one-step generative model (i.e., Pixel MeanFlow) to directly predict the purified image from the adversarial image. This paper further uses LoRA to make the pretrained Pixel Meanflow adaptable to adversarial purification. Experiments show that PMF-Purify achieves SOTA performance across 11 attacks.

Critical strengths
  1. Strong empirical results. PMF-Purify can achieve SOTA purification accuracy with little latency compared to baseline methods (although the robustness evaluation itself is not convicing enough, as discussed in critical weakness 2). These results demonstrate that Pixel MeanFlow (pMF) is a promising backbone for practical adversarial purification (although it is unclear why pMF is particularly suitable, as discussed in critical weakness 1).
Critical weaknesses
  1. The motivation of why pMF is specifically used remains unclear to me (please correct me if I have overlooked something). The motivation of why one-step diffusion is needed is clearly discussed in the paper. However, I would be more interested in understanding why pMF is particularly suitable for adversarial purification, compared to existing one-step generative models. Without a clearer justification, in my point of view, PMF-Purify appears to be an application of an existing generative model to adversarial purification, rather than a fundamentally new purification method. For example, as shown in table 4, base pMF can already do a good job in purifying adversarial examples, making the unique technical contributions of PMF-Purify incremental.
  2. The robustness evaluation is not convincing. All the white-box attacks should optimize the loss by directly backpropagating through the entire purification pipeline (i.e., pMF + downstream classifier) given that PMF-Purify employs a differentiable one-step generative model. To this end, it remains unclear why BPDA is preferred over direct gradient-based attacks. Furthermore, given the Gaussian noise injection, the robustness against PGD+EOT (Lee & Kim, ICCV 2023) should also be evaluated.
Questions for authors

Please refer to my critical weaknesses. I am open to discussions and happy to increase my score if my concerns can be well addressed.

Reviewer_RSAUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PMF-Purify, a method that uses pixel space mean flow for adversarial purification. Because the paper uses mean flow purification, the method can use 1 NFE to purify adversarial inputs back to clean-manifold efficiently. To further improve adversarial purification performance, they propose using LoRA finetuning to push adversarial landscape back to clean data manifold space. They show their methods by using ImageNet and CIFAR-10 datasets, compared with different baselines.

Critical strengths

1- Because they use mean flow, which can generate images in one step, their proposed method is fast and efficient taking only a few milliseconds for purification 2- The paper evaluates their results on high dimensional ImageNET 1K dataset 3- Parameter efficient LoRA finetuning further improves performance by restoring perturbed inputs back to natural image space

Critical weaknesses

1- The main weakness of the paper is the evaluation. The higher results shown in the paper are mainly because of weaker evaluations and not because of the inherent strength of the proposed method. They’re not using real adaptive evaluation that passes through the purification. The only white-box attack uses BPDA-PGD which is also a weaker form of white-box attack as it approximates gradients through the purification with an identity function. I ask the authors to rerun all their experiments with real PGD-EoT and AutoAttack-EoT attacks where gradients pass through the purifier back to the input. Look at [1,2,3] for more on robust evaluation. In addition, since this defense only uses one step for purification, it makes robust white-box evaluation very easy and no need to use BPDA as other diffusion purification works that do have a long chain of purification.

2- Some recent baselines, like purification (AAOpt [3]) and attacks on diffusion purification (DiffBreak[1]), and others are missing from the paper.

3- The novelty of the work is very limited. There are already works very similar to this in the latent space [4, 5]. What differentiates this work mainly is that it’s in the pixel space and using mean flow instead of consistency or other form of distillations.

References:

1-cKassis, Andre, Urs Hengartner, and Yaoliang Yu. "DiffBreak: Is Diffusion-Based Purification Robust?." Advances in Neural Information Processing Systems 38 (2026): 20319-20362.

2- Lee, Minjong, and Dongwoo Kim. "Robust evaluation of diffusion-based adversarial purification." In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 134-144. IEEE, 2023.

3- Wesego, Daniel, and Pedram Rooshenas. "Adversary aware optimization for robust defense." Advances in Neural Information Processing Systems 38 (2026): 38556-38582.

4- Huang, Chihan, Belal Alsinglawi, and Islam Al-Qudah. "DBLP: noise bridge consistency distillation for efficient and reliable adversarial purification." Computers and Electrical Engineering 140 (2026): 111503.

5- Lei, Chun Tong, Hon Ming Yam, Zhongliang Guo, Yifei Qian, and Chun Pong Lau. "Instant adversarial purification with adversarial consistency distillation." In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 24331-24340. IEEE, 2025.

Questions for authors

How does your algorithm perform on attacks that pass through the defense with PGD-EOT & AA-EOT?

How does your algorithm perform compared to AP baselines from [3]?

Could you add additional attacks from [1] with attacks passing through your defense and report how the proposed defense performs?

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

Related papers

Discussion (0)

Sign in to comment.