Shake-It-Off: Jailbreaking Black-Box Large Language Models by Shaking Off Objectionable Semantics
Keywords: Jailbreaking Attacks, Large Language Models
TL;DR: We propose a jailbreaking attack algorithm that requires only API access to the victim model and achieves a better success rate and 10x faster speed than previous methods.
Abstract: Large language models (LLMs) are vulnerable to jailbreaking attacks (Zou et al., 2023; Liu et al., 2024), in which attackers use adversarially designed prompts to bypass the model’s safeguard and force the model to generate objectionable content. The present paper studies jailbreaking attacks from a red team’s viewpoint and proposes a novel black-box attack method, called Shake-It-Off (SHAKE), that only requires the response generated by the victim model. Given objective query $T_{obj}$, our method iteratively shakes off the objectionable semantics of $T_{obj}$, making it gradually approximates a pre-defied decontaminated query $T_{dec}$. We conduct extensive experiments on multiple baseline methods and victim LLMs. The experimental results show that SHAKE outperforms the baseline methods in attack success rates while requiring much less running time and access to the victim model.
Primary Area: foundation or frontier models, including LLMs
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: 724
Loading