Stable Diffusion is Unstable

Recently, text-to-image models have been thriving. Despite their powerful generative capacity, our research has uncovered a lack of robustness in this generation process. Specifically, the introduction of small perturbations to the text prompts can result in the blending of primary subjects with other categories or their complete disappearance in the generated images. In this paper, we propose Auto-attack on Text-to-image Models (ATM), a gradient-based approach, to effectively and efficiently generate such perturbations. By learning a Gumbel Softmax distribution, we can make the discrete process of word replacement or extension continuous, thus ensuring the differentiability of the perturbation generation. Once the distribution is learned, ATM can sample multiple attack samples simultaneously. These attack samples can prevent the generative model from generating the desired subjects without compromising image quality. ATM has achieved a 91.1% success rate in short-text attacks and an 81.2% success rate in long-text attacks. Further empirical analysis revealed four attack patterns based on: 1) the variability in generation speed, 2) the similarity of coarse-grained characteristics, 3) the polysemy of words, and 4) the positioning of words.

Paper

Similar papers

Peer review

Reviewer L6BS6/10 · confidence 4/52023-06-25

Summary

This paper finds some vulnerabilities of stable diffusion model, and proposes an auto-attack model to generate attack prompts.

Strengths

This paper is well written and is easy to understand. This paper discusses some vulnerabilities of the stable diffusion model, and does many experiments to verify. The methodology illustration is clear and easy to follow.

Weaknesses

The purpose and motivation of ATM is not well explained. Authors have find some vulnerabilities of the stable diffusion model, but the motivation of ATM is not correlated well with them. It would be better to connect the ATM and the aforementioned vulnerabilities.

Questions

Are there some related work in this field? I would be better to mention some related work, and discuss or compare them in your experiments. When describing the Variability in Generation Speed, the definition of speed is little confusing. What does the speed value represent? How to understand it?

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

This paper proposes an attack model, but lack of discussion of defense model. It would be better to discuss them together.

Reviewer czYq6/10 · confidence 4/52023-06-29

Summary

The paper introduces Auto-attack on Text-to-image Models (ATM), a method to efficiently generate attack prompts that closely resemble clean prompts. The method modifies text prompts by replacing or extending words, using a Gumbel Softmax distribution for differentiability. It further applies a binary mask to preserve the desired object noun and imposes fluency and similarity constraints to ensure similarity. The authors identify four distinct attack patterns through this attack method. The method utilizes Stable Diffusion as the target model, allowing for white-box attacks, and demonstrates the transferability of these attacks to other generative models (black-box attacks).

Strengths

1. The authors introduce a novel method to generate successful attack prompts in text-to-image generation pipelines. It can aid in vulnerability investigation. The method enables the identification of a broader range of attack patterns, prompting further research in both attack and defense mechanisms. Therefore, it can enhance security in the industry. 2. Related to the previous point, the authors themselves employ their method to uncover four distinct attack patterns, which are highly enlightening. This convincingly demonstrates the significant importance of this approach in studying the vulnerability of generative models and enhancing their robustness. 3. In their experiments, the proposed method demonstrates a high success rate in white-box attacks. It also exhibits excellent transferability to other generative models, e.g. DALL-E 2 and Midjourney. This proves that the method has a broad range of applications and can be used for various generative models, not just limited to Stable Diffusion.

Weaknesses

1. Two types of modifications, namely replacing and extending, are considered. It can be confusing to determine when to use each type. How do they determine whether to replace or extend a word? 2. Most of the analyses are solid. However, in Section 3.1, the authors mention embedding two prompts c_1 and c_2 separately to "eliminate the additional impact of all possible extraneous factors". It is unclear what specific impact they want to eliminate and how embedding prompts separately can achieve this. 3. The authors mention the use of the method to design defensive strategies but do not discuss how to utilize it. Is it possible to provide an example? 4. A minor issue: In Fig. 6, the steps should range from 49 to 0 instead of 1 to 50. This is because in the reverse diffusion process, T represents noise, and 0 represents the generated image (similar to the steps shown in Fig. 4).

Questions

Please address questions mentioned in "Weaknesses".

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

There are no potential negative societal impacts.

Reviewer adct7/10 · confidence 4/52023-07-03

Summary

This paper proposes an adversarial attack against text-to-image models that can generate adversarial prompts to prevent the stable diffusion models from generating the desired subjects. The attack is gradient-based by utilizing the Gumbel Softmax to make the work embedding differentiable. Then, the authors provide a comprehensive analysis of the vulnerabilities of the stable diffusion models.

Strengths

1. This paper is well-written and well-organized. 2. The proposed automatic attack is effective and convenient for implementation. 3. The authors provide interesting and inspiring analyses of the vulnerabilities of the generative models. They found the differences in the generation speed, the similarity of coarse-grained subjects in the prompt, and the polysemy are the important factors in the robustness of the generative models. This analysis is definitely beneficial to the future study of the robustness of generative models. 4. The authors show that the proposed attack is transferable as well.

Weaknesses

1. Although the constraint in the attack seems to be novel, the attack objective is similar to the previous proposed C&W attacks.

Questions

1. What are the computation consumption and requirement to conduct the white-box attack against the stable diffusion? 2. The attack success rate could be affected by the performance of the network. For example, the network could misclassify a subject even if the subject is correctly generated by the diffusion model. Could you discuss how to mitigate this issue during the evaluation procedure? 3. It seems the attack will make the FID score higher and IS score lower. Does it indicate a drawback of the attack? I am confused about whether we need to maintain the FID and IS score during the attack. Could the authors provide some explanations?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, 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

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

None.

Reviewer kKNr7/10 · confidence 4/52023-07-06

Summary

In this paper, the authors use Gumble softmax as well as the gradient based method to learn the distribution of an attack text prompt, defined as an attack text prompt that enables a text-to-image generation model to generate images that do not match the text description without changing the category keywords in the clean text prompt, for exploring the vulnerability of stable diffusion model. Based on the failure cases generated by the algorithm, the authors place the possible reasons for the failure of the generation into four main categories.

Strengths

1. This algorithm able to automatically finds the prompt that causes the text to image generation model to fail, offering the possibility to explore the vulnerability of the stable diffusion model systematically. 2. Various patterns of generative failure have been identified based on a large number of generative failure cases, and a number of reasonable experiments have been designed to further support these conjectures. 3. This algorithm able to learn multiple attack text prompts based on a clean prompt, further boosting the number of generated failure cases.

Weaknesses

1. It seems that in pattern 1 only the speed of generation of different categories under the same random seed was explored, would the speed of generation of the same category with different random seeds also be different? 2. In the section on quantitative experiments, what is the classification accuracy of CLIP for clean text prompts? 3. How do you handle the gradient of 50 steps in the stable diffusion model? 4. Most contents in this paper are clear and detailed, but the setting in the quantitative experiments section lacks some details, e.g. it does not explain how the random algorithm changed the text prompt, which Casual Language Model was used, and which version of Clip was used.

Questions

See weakness

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, 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

3 good

Presentation

3 good

Contribution

3 good

Limitations

None

Reviewer adct2023-08-14

Solved my concerns

Thanks for your replies. I appreciate that the authors used a voting mechanism to eliminate the impact of the performance of a single classifier, which definitely makes the attack success rate more convincible. Other concerns are well solved as well. Therefore, I still lean towards Acceptance.

Reviewer czYq2023-08-15

post-rebuttal comments

The authors have addressed my concerns. It is a good paper. I vote for acceptance.

Area Chair W7kP2023-08-19

Discussions are required for the submission

Dear all reviewers, Thank you very much for your great efforts in reviewing the referred submission. Now the authors have provided responses regarding your concerns. Would you please read the authors response and see whether your concerns have been addressed or not. You are welcome to raise further concerns if necessary so that the authors can respond to them timely. Your great service would be very important for the community to make final decisions. Best regards, Your AC

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC