More Harmful, Less noticeable: Learning Adversarial Null-Text Embeddings for Inconspicuous Attack

Weihang Ran, Mingdeng Cao, Wei Yuan, Isao Echizen, Yinqiang Zheng

Keywords: adversarial attack, diffusion model, image editing

Abstract: Adversarial examples, which are artificially crafted data intended to disrupt the output of deep learning models, present a new round of challenges to the stability and security of artificial intelligence technology. Unrestricted adversarial examples, obtained by modifying the semantic elements of images, have the characteristics of being natural and semantically meaningful. However, previous methods either significantly altered the image's color or content, or blurred visual details (such as text or geometric designs), making the generated adversarial examples easily detectable by the human eye. In this paper, we propose a method to generate highly natural adversarial examples based on stable diffusion. This is achieved by introducing adversarial loss during the image reconstruction process to perturb cross-attention mechanism. To further enhance image quality, we introduce perceptual loss into the adversarial attack process for the first time. Extensive experiments and visualizations demonstrate the effectiveness of our proposed method. Compared to the current state-of-the-art methods, our approach not only improves the adversarial transferability by an average of 12.59-50.3% but also significantly enhances image quality. Code will be publicly available.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1115