AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial Attacks

Imperceptible adversarial attacks aim to fool DNNs by adding imperceptible perturbation to the input data. Previous methods typically improve the imperceptibility of attacks by integrating common attack paradigms with specifically designed perception-based losses or the capabilities of generative models. In this paper, we propose Adversarial Attacks in Diffusion (AdvAD), a novel modeling framework distinct from existing attack paradigms. AdvAD innovatively conceptualizes attacking as a non-parametric diffusion process by theoretically exploring basic modeling approach rather than using the denoising or generation abilities of regular diffusion models requiring neural networks. At each step, much subtler yet effective adversarial guidance is crafted using only the attacked model without any additional network, which gradually leads the end of diffusion process from the original image to a desired imperceptible adversarial example. Grounded in a solid theoretical foundation of the proposed non-parametric diffusion process, AdvAD achieves high attack efficacy and imperceptibility with intrinsically lower overall perturbation strength. Additionally, an enhanced version AdvAD-X is proposed to evaluate the extreme of our novel framework under an ideal scenario. Extensive experiments demonstrate the effectiveness of the proposed AdvAD and AdvAD-X. Compared with state-of-the-art imperceptible attacks, AdvAD achieves an average of 99.9$\%$ (+17.3$\%$) ASR with 1.34 (-0.97) $l_2$ distance, 49.74 (+4.76) PSNR and 0.9971 (+0.0043) SSIM against four prevalent DNNs with three different architectures on the ImageNet-compatible dataset. Code is available at https://github.com/XianguiKang/AdvAD.

Paper

References (56)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ytMn4/10 · confidence 4/52024-07-03

Summary

In this paper, a diffusion-based adversarial attack method is proposed and extensive experiments were conducted to show the effectiveness of the method.

Strengths

1. the paper is well-written and easy to read. 2. both adversarial attack and diffusion model are hot topics. 3. both Linf and L2 results are given in the experiment. 4. An ablation study on T is provided 5. The experiment's results are convincing.

Weaknesses

1. because this paper proposed a new algorithm, at least a pseudo code should be provided to show how the algorithm works. 2. In the table1, it shows that the performance of the model measured in FID sometimes does not match that in other metrics measuring similarity, e.g., psnr, ssim. For example, in the experiment on visionmemba, the FID of the proposed method is not the smallest, as that in other similar metrics. Can the author give some explanation about this?

Questions

same

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

same

Authorsrebuttal2024-08-13

Respectful Request for Further Discussion

Dear Reviewer ytMn, We sincerely thank you for your valuable feedback and the time you have taken to process our submission. As the discussion phase is nearing its end, we respectfully ask for your help once again to review our responses and let us know if they address your concerns. Following your valuable suggestions, we have included the pseudo code for our AdvAD and AdvAD-X in the attached PDF of the global author rebuttal, and it has also been incorporated into the revised version to further enhance clarity. Additionally, we have provided a detailed explanation and corresponding experimental verification for the FID results, which you expressed concerns about. Please also let us know if you have any further questions about this paper. We have made every effort to enhance our work based on your insightful comments, and we would deeply appreciate it if you could further support us! Best regards, The authors

Reviewer nUXD5/10 · confidence 3/52024-07-03

Summary

This work proposes a novel adversarial attack framework called Adversarial Attacks in Diffusion (AdvAD). Unlike prior methods that rely on generative models or specific loss functions, AdvAD formulates attacking as a non-parametric diffusion process. This approach theoretically explores a fundamental modeling strategy instead of leveraging the denoising or generation capabilities of diffusion models with neural networks. AdvAD iteratively refines the attack by crafting subtle adversarial guidance based solely on the targeted DNN, without requiring any additional network. This process progressively transforms the original image into an imperceptible adversarial example.

Strengths

1. The concept of leveraging the core principles of diffusion models to craft adversarial examples is intriguing. 2. The writing style in the Abstract and Introduction sections is engaging. 3. The effectiveness of AdvAD is validated through experiments.

Weaknesses

While I have carefully reviewed the work several times, some uncertainties remain. These are addressed in the following "Questions" section. Due to these uncertainties, the initial recommendation is a borderline decision, which can be revised upwards or downwards based on the authors' response.

Questions

1. The paper claims that AdvAD achieves superior imperceptibility compared to other methods. However, after multiple readings, the justification for this remains unclear. While the text mentions "much subtler yet effective adversarial guidance at each step", "grounded on diffusion model's theoretical foundation", and various diffusion-related formulations, it lacks a clear explanation (or theoretical verification) of how using non-parametric diffusion leads to attacks with lower perturbations. 2. Based on my understanding, AdvAD shares similarities with traditional attacks like PGD, but with a fixed number of optimization iterations (controlled by the total diffusion model timesteps T) and a gradually decaying learning rate (controlled by the $α_t$ value). I hypothesize that the decayed learning rate is a key factor contributing to imperceptibility. The learning rate starts low (at the beginning of T steps) and gradually decreases to 0 (at the end of 0 steps). This suggests that small perturbations are added first, followed by even smaller ones, potentially leading to adversarial examples closer to the classification boundary and thus exhibiting higher imperceptibility. Could the authors comment on the validity of this hypothesis? 3. Following the previous question, to isolate the impact of the decayed learning rate strategy, it would be beneficial to include comparative experiments. These experiments would directly apply a small learning rate initialization and a decayed learning rate schedule to a traditional attack algorithm. If this approach achieves similar performance to AdvAD, it would suggest that the emphasis on diffusion model theory might not be necessary. If not, the authors should highlight the specific benefits of the diffusion model mechanism that contribute to imperceptibility. 4. Does the selection of the initial Gaussian noise sample $ε_0$ affect the performance of AdvAD? 5. In Table 1, is the FID metric computed between the perturbed images and the "imagenet-compatible dataset"? If the goal is to assess the realism of the adversarial examples compared to natural images, a more comprehensive evaluation would be to compare them with the raw ImageNet images. The "imagenet-compatible" dataset, with only 1000 images, might not adequately capture the entire distribution of natural images. 6. Line 248 states that "the optimizer usually cannot find the global optimal solution, and optimization-based methods tend to show sub-optimal ASR." How can AdvAD achieve better ASR compared to these optimization-based methods? 7. In Table 2, which classifier was used to generate the adversarial samples? The robustness of the adversarially trained model Inc-V3_{adv} and the ensemble model Inc-V3_{ens4} appears very low (almost 100% ASR). Additionally, were Res-50, Swin-B, and ConvNeXt-B clean classifiers or adversarially trained? 8. Line 292 suggests that "a larger T exhibits better imperceptibility, while a smaller T implies stronger black-box transferability." Can the authors elaborate on this trade-off? 9. There are a few minor writing errors throughout the paper, such as "a' adversarial" (Line 140), "fistly" (Line 144), and "an' scheme" (Line 191). 10. There are specific questions regarding the proofs of theorems and propositions in the Appendix: (1) In the second row equality of Eq. 6, shouldn’t $\epsilon_T$ be $\epsilon_{T+1}$ in the first term on the numerator of the first term? (2) In Line 546, it Is clear that $\epsilon_{T+1}$ satisfies the inequality as $\epsilon_{T+1}$ equals $\epsilon_0$, but how does this hold for $\epsilon_{T}$ as well? (3) Eq. 20, third row inequality: How is this inequality derived?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, limitations are identified in the paper.

Reviewer jCt24/10 · confidence 4/52024-07-10

Summary

This paper proposes a new method, AdvAD, for generating imperceptible adversarial attacks against deep neural networks (DNNs). AdvAD is based on a novel non-parametric diffusion process, which initializes a fixed diffusion noise and then manipulates it at each step using adversarial guidance crafted by two modules: Attacked Model Guidance (AMG) and Pixel-level Constraint (PC). This process gradually leads the image from its original distribution to a desired adversarial distribution. The paper also introduces an enhanced version, AdvAD-X, which aims to achieve extreme performance under an ideal scenario. Extensive experiments demonstrate the effectiveness of both AdvAD and AdvAD-X in terms of attack success rate, imperceptibility, and robustness compared to state-of-the-art methods.

Strengths

The paper's main strength lies in its novel approach to generating imperceptible adversarial attacks. The use of a non-parametric diffusion process is innovative and shows promising results in terms of attack effectiveness and imperceptibility. The paper is well-written and easy to follow, with clear explanations of the proposed method and experimental setup.

Weaknesses

The paper has several weaknesses. The technical novelty is limited, as the core idea of using diffusion models for adversarial attacks has been explored in prior work (e.g., DiffAttack, ACA). The paper lacks a clear comparison with these existing diffusion-based attack methods, making it difficult to assess the incremental contribution of AdvAD. Another thing is how much cost is it really to generate adversarial examples this way. I mean, one important use of generating adversarial examples is for adversarial training. So if it becomes really expensive to generate adversarial examples, such a use wouldn't be really meaningful. Then what is the use of these adversarial examples really?

Questions

What are the specific differences between AdvAD and existing diffusion-based attack methods like DiffAttack and ACA? A detailed comparison would help clarify the novelty of the proposed approach. What are the potential defensive measures against AdvAD, and how can they mitigate the threat posed by this attack? How does the computational complexity of AdvAD compare to other state-of-the-art methods, especially for large-scale datasets or models?

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A

Authorsrebuttal2024-08-13

Respectful Request for Further Discussion

Dear Reviewer jCt2, We sincerely thank you for your valuable feedback and the time you have taken to process our submission. As the discussion phase is nearing its end, we respectfully ask for your help once again to review our responses and let us know if they address your concerns. In brief, we have provided a detailed point-by-point comparison with other diffusion-based methods from various perspectives to further complement and highlight the significant innovations and contributions of our paper, in addition to those you kindly recognized in the Strengths section. Moreover, we have elaborated on the computational complexity and potential defense methods following your constructive suggestions. It is our pleasure to receive your feedback. Please also let us know if you have any further questions about this paper. We have made every effort to enhance our work based on your insightful comments, and we would deeply appreciate it if you could further support us! Best regards, The authors

Reviewer ygiL6/10 · confidence 4/52024-07-12

Summary

In this paper, the authors propose the Adversarial Attack in Diffusion method called AdvAD, which crafts imperceptible perturbations from the model perspective without the need for additional networks. During the non-parametric diffusion process, the proposed AdvAD method introduces Attacked Model Guidance (AMG) and Pixel-level Constraint (PC) modules to ensure both imperceptibility and transferability of the adversarial examples. Moreover, an enhanced version AdvAD-X is further introduced to achieve better imperceptibility and lower computational complexty. Experiments conducted on CNN, ViT, and Mamba architectures using the ImageNet benchmark dataset demonstrate that the proposed AdvAD method achieves superior performance compared to baseline methods in terms of attack success rate, imperceptibility, and robustness.

Strengths

1. The proposed AdvAD method reformulates the generation of imperceptible adversarial examples as a non-parametric diffusion process, effectively eliminating the negative effects caused by the uncertainty of generative models. 2. The proposed AMG and PC modules inject adversarial guidance progressively at each step without the need for additional networks, and they have a solid theoretical guarantee that the error caused by approximation is strictly bounded. 3. Experimental results verify the effectiveness of the proposed AdvAD method, which achieves higher attack success rate, lower perturbation strength, and better image quality.

Weaknesses

1. In section 3.4, the details of AdvAD-X are not explained clearly. The exact step of adopting original diffusion noise in the DGI strategy is not discussed. Additionally, in the CA strategy, the definition of non-critical image regions is not introduced. 2. In section 4.4, the transferability experiment lacks comprehensiveness. As shown in Table 3, the proposed AdvAD method is only compared to a limited number of baseline methods. 3. In Table 3, with a limited budget of updating iterations, the proposed AdvAD (T=10) performs very similarly to the classic PGD method when using Mob-V2 as a surrogate model, improving the black-box ASR by less than 1%. 4. The experimental settings for Table 4 are not fully introduced, such as the surrogate model used for crafting adversarial examples.

Questions

1. How are non-critical image regions defined for AdvAD-X? Is this definition universally applicable to all images, or is it adjusted for each image separately? 2. Why were other attack methods (such as NCF, ACA, Diff-Attack, Diff-PGD, etc.) not considered in the transferability experiment shown in Table 3? 3. As only a few values (e.g., 10, 100, 1000) are discussed, how can the optimal value of the step $T$ be chosen to ensure both imperceptibility and transferability in practice?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The proposed AdvAD-X method is only suitable for ideal scenarios where the crafted adversarial example is directly input to DNNs without undergoing quantization.

Reviewer K1mG6/10 · confidence 3/52024-07-22

Summary

This paper studies a task about generating the imperceptible adversarial noise using the diffusion model. The proposed method theoretically models the attack process as a non-parametric diffusion process. Extensive experiments demonstrate the effectives of the proposed method.

Strengths

(1) The writing is good and easy to follow. (2) The technique novelty is good, which studies the imperceptible adversarial perturbation generation as a non-parametric diffusion process. The behind theory is solid. (3) Authors conduct extensive experiments and compare many baselines to show the effectiveness of the proposed method.

Weaknesses

(1 ) The adopted defense methods are not strong, lacking some strong baselines, like the random smoothing. [1] Cohen, Jeremy, Elan Rosenfeld, and Zico Kolter. "Certified adversarial robustness via randomized smoothing." international conference on machine learning. PMLR, 2019. (2) Could the generated adversarial examples used for black-box (or transfer-based) attack settings? What's the limitation?

Questions

see weakness

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

see weakness

Reviewer nUXD2024-08-11

Thanks for the response.

After reviewing the authors' response and considering the innovative nature of the attack they have implemented, I have decided to raise my score to 5.

Authorsrebuttal2024-08-11

Thank you for the response. We sincerely appreciate your valuable comments and are grateful for the recognition of our work.

Authorsrebuttal2024-08-12

Experimental verification regarding the previous response to the comment on FID metric (Weakness 2).

To further validate the explanation provided above and offer a better understanding, we directly calculate the cosine similarity of the extracted global features between all adversarial examples and their corresponding original images. The average results for the two backbones are shown in the last two rows of the following table, respectively. | Metric | ConvNeXt | | VisionMamba | | |------------------------------------|----------|------------|-------------|------------| | | **SSAH** | **AdvAD** (ours) | **SSAH** | **AdvAD** (ours) | | **ASR (%) $\\uparrow$** | 84.6 | 100.0 | 49.8 | 99.7 | | **$l_2$ $\\downarrow$** | 2.24 | 1.49 | 1.95 | 1.62 | | **FID $\\downarrow$** | 3.04 | 5.07 | 2.08 | 3.67 | | **Avg. Feature Similarity (Attacked Model)** | 0.5745 | 0.4184 | 0.5424 | 0.2321 | | **Avg. Feature Similarity (Inception)** | 0.9908 | 0.9845 | 0.9938 | 0.9888 | It can be more easily observed that, consistent with our previous explanation, the lower FID results of the SSAH attack are due to its ineffectiveness in attacking the ConvNeXt and VisionMamba backbones. As shown in the sixth row of the table, despite the higher strength of the perturbation injected into the adversarial examples (as indicated by the $l_2$​ metric), the feature similarity of SSAH remains higher than ours. This suggests that the adversarial perturbation crafted by SSAH is more difficult to impact the high-dimensional features on which the model's final classification decision relies, leading to its lower ASR. For the Inception model, which is used to calculate the FID metric, SSAH's feature similarity is anomalously further amplified to above 0.99 (the last row), indicating that the adversarial perturbation crafted by SSAH has almost no effect on the high-dimensional features of the clean images for the Inception model, resulting in its lower FID results. However, when all attacks are performing correctly, the magnitude of changes in image-space and feature-space can be considered as positively correlated in the scenario of imperceptible attacks. In this case, benefiting from the inherently lower perturbation strength of our modeling approach, AdvAD consistently achieves the best FID score when attacking the other models, as shown in Table 1 and Table 5 (Appendix).

Reviewer ygiL2024-08-12

Thanks for the authors’ responses. My concerns have been properly addressed.

Reviewer K1mG2024-08-13

Thanks for your response. It addresses most of my concerns. Hope authors could include these discussions into the revision.

Authorsrebuttal2024-08-13

Thank you for the response. We appreciate your valuable review and we will incorporate the discussions to the revised version accordingly.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC