Some text is shown in its original language while the English translation is being prepared.
BackSource arXiv
Md Rafi Ur Rashid, MD Sadik Hossain Shanto, Vishnu Asutosh Dasu, Shagufta Mehnaz
Indexed by this site 2026-10-12 14:28 2 hr ago
Originally published Feb 9
Selected 2026-10-12 14:38
Category Attacks arXiv · arXiv:2602.08136 · Md Rafi Ur Rashid, MD Sadik Hossain Shanto, Vishnu Asutosh Dasu, Shagufta Mehnaz · Indexed by this site 2026-10-12 14:28 · Originally publishedFeb 9 SelectedAttention53 视觉语言模型对拆图有害输入攻击的鲁棒性 Robustness of Vision Language Models Against Split-Image Harmful Input Attacks
AI reading guide
论文发现视觉语言模型的安全对齐只针对完整图像,当有害语义分散到拆分后的多张图像时,模型往往无法识别并拒绝。作者提出拆图视觉越狱攻击 SIVA,包含 Direct、Adaptive、Transfer 三种独立威胁模型设置,在 GPT-5.2、Gemini 2.5 Flash、Qwen3-VL 30B 和 Llama-4-Scout-17B-16E-Instruct 四个模型、三个越狱数据集上评估,最强的 Transfer-SIVA 借助对抗知识蒸馏在完全黑盒设置下比基线最高提升约 44% 的迁移成功率。论文还提出一种复用偏好标签的 DPO 变体,在安全对齐阶段纳入拆图样本以缓解该漏洞。
Deep reading Generating
Why it was selected
论文指出视觉安全对齐只覆盖完整图像,拆分输入中的分散危害线索未被拒绝,并给出可在对齐层补上的做法。
Read original arxiv.org Why it was selected SelectedAttention53
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文指出视觉安全对齐只覆盖完整图像,拆分输入中的分散危害线索未被拒绝,并给出可在对齐层补上的做法。