Towards good practice in boosting the targeted adversarial attack
Keywords: adversariak attack
Abstract: By accessing only the surrogate model, attackers can craft adversarial perturbations to fool black-box victim models into misclassifying a given image into the target class. However, the misalignment between surrogate models and victim models raises concerns about defining what constitutes a successful targeted attack in a black-box setting. In our work, we empirically identify that the vision-language foundation model CLIP is a natural good indicator to evaluate a good transferable targeted attacks. We find that a successful transferable targeted attack not only confuse the model on the vision modality towards the target class, but also fool the model on the text modality between the original class and target class. Motivated by this finding, we propose a simple yet effective regularization term to boost the existing transferable targeted attacks. We also revisit the feature-based attacks, and propose to boost the performance by enhancing the fine-grained features. Extensive experiments on the ImageNet-1k dataset demonstrate the effectiveness of our proposed methods. We hope our finding can motivate future research on the understanding of targeted attacks and develop more powerful techniques.
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: 845
Loading