acceptodds
Under review as a conference paper at ICLR 2027

Fast Preemptive Robustification: High-Frequency Response Anti-Aligns Shared Vulnerability

Abstract

Adversarial attacks can readily compromise deep neural networks (DNNs). In particular, transferable attacks (TAs) exploit the shared vulnerabilities among DNNs, enabling perturbations crafted on surrogates to transfer to unseen models. Training-time and post-attack defenses have been extensively studied for combating TAs. Orthogonal to these approaches, preemptive robustification (PR) has emerged as a pre-attack defense that enhances the robustness of benign samples by superimposing protective variations before attacks. Despite its promise, PR remains underexplored and faces several important limitations. First, dependence on well-trained surrogate classifiers limits applicability, as surrogates are task-specific and may even be unavailable in some practical settings. Moreover, the required iterative optimization or dedicated PR generator training incurs substantial costs. Third, the variations generated by existing methods lack human interpretability. To address these, we seek an efficient PR that is surrogate-free, optimization-free, training-free, and human-interpretable. Intriguingly, we discover a numerical correlation between the shared vulnerabilities of DNNs and Laplacian responses, with their cosine similarity being significantly negative. This indicates that negated high-frequency response constitutes an important component of shared vulnerabilities. Consequently, strengthening Laplacian responses counteracts this component, improving resistance to TAs. Building upon this insight, we propose **F**ast **P**reemptive **R**obustification (**FPR**), which runs efficiently on commodity CPUs, enabling real-time PR without dedicated GPUs. Despite its simplicity, FPR provides substantial robustness gains across diverse attack settings. Specifically, FPR reduces the attack success rate (ASR) of untargeted TAs by and that of targeted TAs from to . The code will be released.

Primary area:
interpretability and explainable AI
Keywords:
Computer Vision, Adversarial Machine Learning, Preemptive Robustification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dV2zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Fast Preemptive Robustification (FPR), which runs efficiently on commodity CPUs, enabling real-time PR without dedicated GPUs. Regarding the current PR, it remains underexplored and faces several important limitations. First, dependence on well-trained surrogate classifiers limits applicability, as surrogates are task-specific and may even be unavailable in some practical settings. Moreover, the required iterative optimization or dedicated PR generator training incurs substantial costs. Third, the variations generated by existing methods lack human interpretability. FPR exploits the observation of Laplacian response to improve the robustness, and the experimental results show that FPR can achieve better performance.

Critical strengths
  1. The research problem is interesting. Preemptive robustification (PR) proactively enhances the robustness of input samples before attacks, rather than relying on model training or post-attack defenses. FPR further introduces a more efficient approach to PR.
  2. The proposed FPR method is simple and computationally efficient, requiring low computational resources.
  3. The experimental evaluation is broad, covering various attack scenarios and experimental settings.
  4. The paper is well-written and easy to follow, and the figures are clear and informative.
Critical weaknesses
  1. The methodological novelty is limited. The proposed method essentially applies an existing technique to PR. Although the research problem is interesting, the novelty of the method itself is not sufficiently compelling.
  2. The lack of evaluation against adaptive attacks may lead to an overestimation of the robustness improvements. FPR is a deterministic and known image transformation with a very simple formulation. If attackers are aware of its mechanism and parameters, they may develop targeted strategies to undermine its effectiveness, such as explicitly optimizing perturbations that counteract the high-frequency enhancement introduced by FPR.
  3. The hyperparameter selection may be challenging in practical scenarios. One key parameter of FPR is the sharpening coefficient α, which appears to be related to the desired protection strength and the attack intensity. However, the attack strength and perturbation magnitude are typically unknown in real-world scenarios, making it difficult to determine an appropriate α. The paper only evaluates different values of α but lacks further discussion or analysis of how this parameter should be selected in practice, particularly whether it can be determined adaptively.
  4. The claimed computational efficiency lacks sufficient experimental validation. Although the paper emphasizes reduced computational costs, it does not provide direct comparisons of computational overhead with existing PR methods.
Questions for authors

See weakness.

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

This paper studies preemptive image robustification to improve images’ resistance to transferable adversarial attacks. The authors observe a negative correlation between the average white-box adversarial perturbations across multiple classifiers and the Laplacian responses of images, and propose Fast Preemptive Robustification (FPR) based on this observation. FPR generates robustified images through Laplacian sharpening, requiring only a single channel-wise convolution without access to classifiers, model training, or iterative optimization. The paper evaluates untargeted, targeted, and ensemble-based transfer attacks using multiple CNN and Transformer models on the NIPS 2017 Adversarial Competition dataset. The authors report that FPR reduces the average attack success rates of single-surrogate untargeted attacks and ensemble-based attacks by approximately 12.7 and 14.7 percentage points, respectively, and lowers the targeted attack success rate from 10.7% to 4.1%. The paper also evaluates changes in clean accuracy, defense effectiveness after JPEG compression, and gains when combined with existing robust models.

Critical strengths
  1. The method is simple and easy to deploy. Equation (9) clearly describes how robustified images are generated without access to downstream classifiers. This makes the approach practically appealing for preemptive defense during image uploading, storage, and transmission.
  2. The evaluation covers a broad range of transfer attacks. Section 4 includes six untargeted, four targeted, and three ensemble-based attacks, using both CNN and Transformer models. Figures 5 and 7 show gains across multiple perturbation budgets and surrogate models, providing substantial empirical support for the method’s effectiveness.
  3. The evaluation considers factors beyond robustness gains. Figure 9 reports clean accuracy, Figure 10 examines JPEG compression applied after sharpening, and Figure 8 explores combinations with existing robust models. These experiments help assess deployment trade-offs.
Critical weaknesses

1 The main transfer-attack experiments lack convergence and parameter sensitivity checks Sections 4.2 and 4.4 use 10-step attacks, whereas the mechanism analysis in Figure 2 runs for up to 1,000 steps. The experiments lack attack parameter sensitivity checks and do not report attack success rates on the surrogate models, making it difficult to determine whether the defense gains stem from increased difficulty in attacking the surrogate or reduced transferability across models. I recommend evaluating different iteration counts, step sizes, and random initializations, and reporting the strongest transfer-attack configuration together with attack success rates on both surrogate and victim models. White-box diagnostics could provide supplementary evidence for the proposed mechanism. 2 Direct defense baselines and quantitative assessment of protective modifications are missing Section 4 lacks direct comparisons with Bi-Level and A⁵, making it difficult to assess FPR’s advantages in defense effectiveness, computational overhead, and image quality. The paper also does not report the distribution of protective perturbation norms, and the limited examples in Figure 4 do not support general claims of imperceptibility or preservation of visual quality. I recommend comparing existing PR methods and simple image enhancement baselines under matched modification magnitudes or perceptual quality, and reporting perturbation norms, perceptual metrics, and enlarged local views. For A⁵, the extent to which its certification setting is comparable to this paper’s transfer-attack evaluation should be clarified. 3 The claim of real-time processing on commodity CPUs lacks measured evidence The paper emphasizes real-time processing on commodity CPUs but reports only an A100-based experimental environment, without CPU latency, throughput, or runtime comparisons with existing PR methods. A single convolution does not by itself establish real-time performance. I recommend reporting measurements on representative CPUs, specifying the hardware model, image resolution, batch size, and operations included in the timing. 4 The signs of the similarity values are inconsistent Lines 283 and 294 on page 6 report values of approximately 0.2 and 0.3, respectively, whereas Figure 2, the abstract, and the argument indicate a negative correlation. These values should be explicitly written as approximately -0.2 and -0.3, or identified as absolute values. 5 Implementation details need to be completed Please specify the exact Laplacian kernel matrix, its sign convention and normalization, boundary padding, the order of sharpening and resizing, and the model weights and preprocessing configurations.

Questions for authors
  1. Are perturbations regenerated from the robustified images for all attacks, with the perturbation budget centered on those images? Please confirm that the implementation is consistent with Equation (5) and clarify what information about the defense is available to the attacker.
  2. How many images are included in the correlation analysis in Figure 2? Are dataset-wide statistics and control results involving random perturbations, clipping, or shuffled image–perturbation pairs available?
  3. Which α values correspond to the reductions of 12.7 and 14.7 percentage points and the decrease from 10.7% to 4.1%, and how are these results aggregated? Are results for different iteration counts and step sizes, along with attack success rates on the surrogate models, available?
  4. How is the Laplacian kernel implemented, and how is α selected? Are CPU latency measurements and comparable results for methods such as Bi-Level available?
Reviewer_Wcg3OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies preemptive robustification: a data owner modifies an image before an attacker can manipulate it, aiming to make later attacks less effective. The focus is on transferable attacks, where perturbations crafted using one classifier also fool other classifiers. The authors observe that averaged attack perturbations point against the image's high-frequency response, and propose to strengthen this response using Laplacian sharpening. The method requires only one 3×3 convolution, without accessing a classifier or optimizing each image. Experiments on the NIPS 2017 Adversarial Competition dataset cover CNNs, transformers, and several transfer attacks. Results report a 12.7 percentage-point reduction in average untargeted attack success, with some loss of clean accuracy.

Critical strengths
  1. The method is simple and easy to apply. A single convolution requires no access to the classifier, no model training, and no optimization for individual images.

  2. The evaluation covers a useful range of transfer attacks, including untargeted, targeted, and ensemble attacks, with both CNN and transformer classifiers and several perturbation budgets.

  3. I appreciate the experiments on clean accuracy, JPEG compression, and combining the method with adversarial training. These help explain how the transformation would behave in practice.

Critical weaknesses
  1. Lack of novelty: pre-processing an image before it reaches an image classifier has been well-studied [1,2]. The motivation is also not novel: there have been extensive defenses motivated by the image’s frequency features [3,4].

  2. Miss important reference. PubDef [5] trains against transfer attacks from public surrogate models, while Ensemble Adversarial Training [6] uses perturbations transferred from other models; neither is cited or discussed.

[1] Countering Adversarial Images Using Input Transformations, ICLR’18

[2] Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks, NDSS’18

[3] SHIELD: Fast, Practical Defense and Vaccination for Deep Learning Using JPEG Compression, KDD’18

[4] Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples, CVPR’19

[5] PubDef: Defending Against Transfer Attacks From Public Models, ICLR’24

[6] Ensemble Adversarial Training: Attacks and Defenses, ICLR’18

Questions for authors

Can FPR protect other attacks like query-based black-box attacks? If it cannot lend protection to worst-case white-box attacks, preprocessing might be able to defend against various black-box attacks, and evaluations there are encouraged, maybe allowing authors to frame the story differently.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.