arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2404.03411cs.LGcs.CLcs.CR
收藏

红队测试GPT-4V:GPT-4V能否抵御单/多模态越狱攻击?

Red Teaming GPT-4V: Are GPT-4V Safe Against Uni/Multi-Modal Jailbreak Attacks?

LMU Munich(慕尼黑大学);Siemens AG(西门子公司);Wuhan University(武汉大学);University of Oxford(牛津大学);Munich Center for Machine Learning (MCML)(慕尼黑机器学习中心)

Shuo Chen, Zhen Han, Bailan He, Zifeng Ding, Wenqian Yu, Philip Torr, Volker Tresp, Jindong Gu

AI总结:

本研究构建了含1445个有害问题的越狱评估数据集,对11个LLM/MLLM进行红队测试,发现GPT-4V等闭源模型更鲁棒,视觉越狱方法迁移性有限。

摘要:

各种越狱攻击已被提出,用于对大型语言模型(LLMs)进行红队测试,并揭示了LLMs脆弱的防护机制。此外,一些方法不局限于文本模态,通过扰动视觉输入将越狱攻击扩展到多模态大型语言模型(MLLMs)。然而,缺乏通用的评估基准使得性能复现和公平比较变得复杂。此外,对闭源最先进(SOTA)模型,尤其是MLLMs(如GPT-4V)的全面评估仍然不足。为解决这些问题,本研究首先构建了一个全面的越狱评估数据集,包含1445个有害问题,覆盖11种不同的安全策略。基于该数据集,对11个不同的LLMs和MLLMs进行了广泛的红队实验,包括SOTA专有模型和开源模型。随后,我们对评估结果进行了深入分析,发现:(1)GPT-4和GPT-4V在抵御越狱攻击方面表现出比开源LLMs和MLLMs更强的鲁棒性。(2)Llama2和Qwen-VL-Chat比其他开源模型更鲁棒。(3)视觉越狱方法的可迁移性相对有限,与文本越狱方法相比。数据集和代码可在https://github.com/chenxshuo/RedTeamingGPT4V找到。

英文摘要:

Various jailbreak attacks have been proposed to red-team Large Language Models (LLMs) and revealed the vulnerable safeguards of LLMs. Besides, some methods are not limited to the textual modality and extend the jailbreak attack to Multimodal Large Language Models (MLLMs) by perturbing the visual input. However, the absence of a universal evaluation benchmark complicates the performance reproduction and fair comparison. Besides, there is a lack of comprehensive evaluation of closed-source state-of-the-art (SOTA) models, especially MLLMs, such as GPT-4V. To address these issues, this work first builds a comprehensive jailbreak evaluation dataset with 1445 harmful questions covering 11 different safety policies. Based on this dataset, extensive red-teaming experiments are conducted on 11 different LLMs and MLLMs, including both SOTA proprietary models and open-source models. We then conduct a deep analysis of the evaluated results and find that (1) GPT4 and GPT-4V demonstrate better robustness against jailbreak attacks compared to open-source LLMs and MLLMs. (2) Llama2 and Qwen-VL-Chat are more robust compared to other open-source models. (3) The transferability of visual jailbreak methods is relatively limited compared to textual jailbreak methods. The dataset and code can be found https://github.com/chenxshuo/RedTeamingGPT4V

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑