EUVIMEDEUVIMED entdecken
Crossref

Visual Adversarial Examples Jailbreak Aligned Large Language Models

Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, Prateek Mittal

Proceedings of the AAAI Conference on Artificial Intelligence · 24.03.2024 · 38 (19) · 21527-21536

DOI: 10.1609/aaai.v38i19.30150 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence, online erschienen am 24.03.2024. Verfasst von Xiangyu Qi und 5 weiteren Autor:innen. Die Arbeit wurde bisher 167-mal zitiert (Crossref).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Warning: this paper contains data, prompts, and model outputs that are offensive in nature. Recently, there has been a surge of interest in integrating vision into Large Language Models (LLMs), exemplified by Visual Language Models (VLMs) such as Flamingo and GPT-4. This paper sheds light on the security and safety implications of this trend. First, we underscore that the continuous and high-dimensional nature of the visual input makes it a weak link against adversarial attacks, representing an expanded attack surface of vision-integrated LLMs. Second, we highlight that the versatility of LLMs also presents visual attackers with a wider array of achievable adversarial objectives, extending the implications of security failures beyond mere misclassification. As an illustration, we present a case study in which we exploit visual adversarial examples to circumvent the safety guardrail of aligned LLMs with integrated vision. Intriguingly, we discover that a single visual adversarial example can universally jailbreak an aligned LLM, compelling it to heed a wide range of harmful instructions (that it otherwise would not) and generate harmful content that transcends the narrow scope of a `few-shot' derogatory corpus initially employed to optimize the adversarial example. Our study underscores the escalating adversarial risks associated with the pursuit of multimodality. Our findings also connect the long-studied adversarial vulnerabilities of neural networks to the nascent field of AI alignment. The presented attack suggests a fundamental adversarial challenge for AI alignment, especially in light of the emerging trend toward multimodality in frontier foundation models.

Zitieren

APA

Qi, X., Huang, K., Panda, A., Henderson, P., Wang, M., & Mittal, P. (2024). Visual Adversarial Examples Jailbreak Aligned Large Language Models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(19), 21527-21536. https://doi.org/10.1609/aaai.v38i19.30150

Vancouver

Qi X, Huang K, Panda A, Henderson P, Wang M, Mittal P. Visual Adversarial Examples Jailbreak Aligned Large Language Models. Proceedings of the AAAI Conference on Artificial Intelligence. 2024;38(19):21527-21536. doi:10.1609/aaai.v38i19.30150

BibTeX
@article{Qi2024,
  author = {Qi, Xiangyu and Huang, Kaixuan and Panda, Ashwinee and Henderson, Peter and Wang, Mengdi and Mittal, Prateek},
  title = {Visual Adversarial Examples Jailbreak Aligned Large Language Models},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year = {2024},
  volume = {38},
  number = {19},
  pages = {21527--21536},
  doi = {10.1609/aaai.v38i19.30150}
}
Kostenlos starten →