Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
通过物理物体对视觉-语言-动作模型发起目标导向后门攻击
[Daily Paper] Goal-oriented Backdoor Attack against Vision-Language-Action Models via Physical Objects
AI summary
研究者提出 GoBA(Goal-oriented Backdoor Attack),通过向机器人环境注入物理世界中的触发物体,静默把 VLA 模型的行为重定向到攻击者指定的目标动作,同时不降低模型在干净输入上的表现。该攻击针对视觉-语言-动作模型,触发条件来自物理环境中的物体。
Read originalfailurefirst.org