BadLIBERO
收藏资源简介:
BadLIBERO数据集是基于LIBERO基准测试构建的,它包含了各种物理触发器和相应的目标导向后门动作。该数据集共有240条数据,用于研究目标导向的后门攻击,旨在使受攻击的视觉语言动作模型在物理触发器存在时执行预定义和目标导向的动作,而在无触发器时表现正常。数据集通过将后门动作轨迹和触发器颜色等因素进行系统分析,为未来的攻击策略设计提供了指导。
The BadLIBERO dataset is built upon the LIBERO benchmark, which encompasses various physical triggers and corresponding goal-directed backdoor actions. Comprising a total of 240 samples, this dataset is developed for research on goal-directed backdoor attacks, aiming to enable compromised vision-language action models to execute predefined and goal-directed actions when physical triggers are present, while maintaining normal behavior in the absence of triggers. By systematically analyzing factors including backdoor action trajectories and trigger colors, the dataset provides guidance for the design of future attack strategies.

- 1通过奥克兰大学 · 2025年



