Beyond Visual Safety: Jailbreaking Multimodal Large Language Models for Harmful Image Generation via Semantic-Agnostic Inputs
超越视觉安全:通过语义无关输入对多模态大语言模型进行有害图像生成的劫持
机构 * State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications(网络与交换技术国家重点实验室,北京邮电大学) ; School of Cyberspace Security, Beijing University of Posts and Telecommunications(网络安全学院,北京邮电大学)
专题命中 可控生成 :image generation(title);分类 cs.CV
AI总结 本文提出BVS框架,通过语义无关输入对多模态大语言模型进行有害图像生成的劫持,揭示其视觉安全边界的脆弱性。