EventPast event
论文提出将不可信内容渲染为图像以防御提示注入
Start here
AI overview
arXiv 论文提出防御方法 Pictionary：多模态大语言模型对以文本形式出现的对抗指令比对图像等非文本通道的同一指令更易服从，据此在不可信内容进入模型前先将其渲染为排版图像或音频，该防御无需训练。随后在 X 上，@florian_tramer 评论称，当 @JieZhang_ETH 提出这个项目时，他开玩笑说自己知道这想法来自本人而非 LLM，因为它太古怪，并称这是赞美，认为在被 AI 垃圾研究淹没时尝试古怪方向比以往更重要。
Generated by AI from reports · Updated 3 days ago
Latest update
张杰谈VLM提示注入防御思路Latest developments
2 developments
- · 1 reports张杰谈VLM提示注入防御思路
- · 1 reports论文提出将不可信内容渲染为图像以防御提示注入
Progress timeline
Oct 2
- X @florian_tramer张杰谈VLM提示注入防御思路
当 @JieZhang_ETH 提出这个项目时，我开玩笑说我知道这想法来自他而不是 LLM，因为它太古怪了。 我这是赞美！ 当我们被 AI 垃圾研究淹没时，偏离常规、尝试古怪的东西比以往任何时候都更重要。
- X@JieZhang_ETH· DiscussionPictionary作者讨论将不可信文本渲染为图片的提示注入防御(opens the original post in a new tab)
Sep 29
- arXivRender Before Reading：把不可信内容渲染成图像以防御提示注入
论文发现多模态大语言模型对以文本形式出现的对抗指令比对图像等非文本通道的同一指令更易服从，并将这一模态差异转化为无需训练的防御方法 Pictionary：在不可信内容进入模型前先渲染为排版图像或音频。
Related discussion
Total 1 items
- Older than 72 hours · Not included in the 72-hour figures
- X@JieZhang_ETH · Jie ZhangPictionary作者讨论将不可信文本渲染为图片的提示注入防御(opens the original post in a new tab)
“We know VLMs mostly learn instruction-following from text, so for them an image is something to describe, not obey.”
Attention trend
New sources per day
- Sep 29New1 sources (1 outlets,0 accounts)
- Sep 30New0 sources (0 outlets,0 accounts)
- Oct 1New0 sources (0 outlets,0 accounts)
- Oct 2New2 sources (1 outlets,1 accounts)
- Oct 3New0 sources (0 outlets,0 accounts)
- Oct 4New0 sources (0 outlets,0 accounts)
- Oct 5New0 sources (0 outlets,0 accounts)
- Oct 6New0 sources (0 outlets,0 accounts)
- Oct 7New0 sources (0 outlets,0 accounts)
- Oct 8New0 sources (0 outlets,0 accounts)
- Oct 9New0 sources (0 outlets,0 accounts)
- Oct 10New0 sources (0 outlets,0 accounts)
- Oct 11New0 sources (0 outlets,0 accounts)
- Oct 12New0 sources (0 outlets,0 accounts)