Fixed Strength Optimization Enhances Adversarial Attacks

Huixin Chen, Dan Hu

Keywords: Adversarial examples, Attack efficiency， Transferability，Perturbation imperceptibility，Distance metric, Multi-step optimization

TL;DR: We propose an optimized adversarial example generation method (FSO) that enhances attack transferability under fixed strength, and introduce a combined norm to improve perturbation imperceptibility.

Abstract: Gradient-based multi-step iteration has been widely used to enhance attack efficiency of adversarial examples. In this work, we propose a $\textit{Fixed Strength Optimization}$ (FSO) method to accelerate the convergence of adversarial examples with a fixed preset attack strength. FSO can be easily combined with existing attack techniques to achieve fast convergence and well-controlled attack strength. We further introduce a combined norm based on $L_{2}$ and $L_{\infty}$ norms to modulate the attacking direction. This combined norm can help to balance the attack strength in the directions of semantic information and noise components in the model gradients on sample data. By incorporating the combined norm into FSO, our numerical experiments show improved attack transferability and high imperceptibility of perturbations.

Primary Area: interpretability and explainable AI

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: 9874