Eugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, Vitaly Shmatikov
using Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMsEugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, Vitaly Shmatikov[Cornell Tech]用图像和声音对多模态LLM进行间接指令注入动机:当前的多模态大型语言模型存在潜在的安全风险,攻击者可以将敌...