VisualDAN:利用视觉驱动的 DAN 指令暴露 VLM 中的漏洞
VisualDAN: Exposing Vulnerabilities in VLMs with Visual-Driven DAN Commands
AI总结:
本文提出 VisualDAN,一种嵌入 DAN 风格指令的对抗性图像,通过在有害语料前添加肯定性前缀训练并转换至文本域,成功绕过 MiniGPT-4 等 VLMs 的安全防护,揭示了基于图像的攻击漏洞并强调亟需鲁棒防御。
摘要:
视觉语言模型(VLMs)因其解释和生成多模态内容的卓越能力而备受关注。然而,保护这些模型免受越狱攻击仍然是一个重大挑战。与纯文本模型不同,VLMs 集成了额外的模态,引入了图像劫持等新型漏洞,这可能操纵模型产生不当或有害的响应。受诸如“Do Anything Now”(DAN)指令等基于文本的越狱启发,本工作提出了 VisualDAN,这是一种嵌入了 DAN 风格指令的单张对抗性图像。具体而言,我们在有害语料前添加肯定性前缀(例如,“Sure, I can provide the guidance you need”),以诱骗模型对恶意查询做出积极回应。然后,对抗性图像基于这些受 DAN 启发的有害文本进行训练,并转换到文本域以引发恶意输出。在 MiniGPT-4、MiniGPT-v2、InstructBLIP 和 LLaVA 等模型上的大量实验表明,VisualDAN 能有效绕过已对齐 VLMs 的安全防护,迫使它们执行严重违反伦理标准的广泛有害指令。我们的结果进一步表明,一旦模型的防御被攻破,即使是少量的有毒内容也能显著放大有害输出。这些发现突显了针对基于图像攻击的鲁棒防御的迫切需求,并为未来 VLMs 的对齐与安全研究提供了关键见解。
英文摘要:
Vision-Language Models (VLMs) have garnered significant attention for their remarkable ability to interpret and generate multimodal content. However, securing these models against jailbreak attacks continues to be a substantial challenge. Unlike text-only models, VLMs integrate additional modalities, introducing novel vulnerabilities such as image hijacking, which can manipulate the model into producing inappropriate or harmful responses. Drawing inspiration from text-based jailbreaks like the "Do Anything Now" (DAN) command, this work introduces VisualDAN, a single adversarial image embedded with DAN-style commands. Specifically, we prepend harmful corpora with affirmative prefixes (e.g., "Sure, I can provide the guidance you need") to trick the model into responding positively to malicious queries. The adversarial image is then trained on these DAN-inspired harmful texts and transformed into the text domain to elicit malicious outputs. Extensive experiments on models such as MiniGPT-4, MiniGPT-v2, InstructBLIP, and LLaVA reveal that VisualDAN effectively bypasses the safeguards of aligned VLMs, forcing them to execute a broad range of harmful instructions that severely violate ethical standards. Our results further demonstrate that even a small amount of toxic content can significantly amplify harmful outputs once the model's defenses are compromised. These findings highlight the urgent need for robust defenses against image-based attacks and offer critical insights for future research into the alignment and security of VLMs.
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。