Some text is shown in its original language while the English translation is being prepared.
BackSource arXiv
Chen Ling, Kai Hu, Hangcheng Liu, Xingshuo Han, Tianwei Zhang, Changhai Ou
Indexed by this site 2026-10-12 14:29 2 hr ago
Originally published Date not given
Selected 2026-10-12 14:42
Category Attacks arXiv · arXiv:2601.17383 · Chen Ling, Kai Hu, Hangcheng Liu, Xingshuo Han, Tianwei Zhang, Changhai Ou · Indexed by this site 2026-10-12 14:29 · Originally publishedDate not given SelectedAttention56 针对大型视觉语言模型的物理提示注入攻击 Physical Prompt Injection Attacks on Large Vision-Language Models
AI reading guide
研究者提出 PPIA,一种黑盒、与用户查询无关的物理提示注入攻击,通过在物理物体上嵌入印刷文本指令来操纵大型视觉语言模型(LVLM)。作者称这是首个物理提示注入攻击,攻击者无需接触模型、输入通道或用户查询,仅通过视觉观察即可实施。方法先用 LLM 生成候选恶意提示词,再以交叉熵衡量其在图像中的可识别性来筛选,并借助 CLIP 的时空注意力图选择部署位置。在 Habitat 和 Embodied City 模拟环境中对 10 个主流 LVLM 测试,视觉问答、任务规划和导航三类任务上的攻击成功率在 70% 到 98% 之间;在搭载摄像头的无人地面车辆上,导航任务成功率超过 80%,在距离、旋转、光照和运动模糊等条件下仍有较好表现。防御方面,严格的 OCR 掩码能最有效降低攻击成功率,但会一并屏蔽任务所需的正常文本。
Deep reading Generating
Why it was selected
论文在 10 个 LVLM 上验证了物理世界视觉提示注入,部署多模态感知系统的团队可据此评估环境文本带来的风险。
Read original arxiv.org Why it was selected SelectedAttention56
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文在 10 个 LVLM 上验证了物理世界视觉提示注入,部署多模态感知系统的团队可据此评估环境文本带来的风险。