Consistent attack: Universal adversarial perturbation on embodied vision navigation

Chengyang Ying, You Qiaoben, Xinning Zhou, Hang Su, Wenbo Ding, Jianyong Ai

Abstract: Highlights•To our knowledge, this is the first attempt to study the robustness of Embodied Vision Navigation against UAP.•We formulate universal perturbation in Embodied agents as δ-MDP and propose two Consistent Attack methods.•We apply our methods to Habitat to show the efficiency of our attack methods.•We empirically demonstrate that existing Embodied Vision Navigation algorithms are vulnerable to our methods.

External IDs:dblp:journals/prl/YingQZSDA23