EIA: Edge-Aware Imperceptible Adversarial Attacks on 3D Point Clouds

Zhensu Wang, Weilong Peng, Le Wang, Zhizhe Wu, Peican Zhu, Keke Tang

Abstract: Adversarial attacks on point clouds are crucial for assessing and improving the adversarial robustness of 3D deep learning models. Existing methods typically apply perturbations to all points on the point cloud using the same strategy. However, compared to flatter regions, human eyes have a higher tolerance for perturbations in areas with more drastic changes, i.e., edges. Based on this consideration, we propose a novel framework named Edge-aware Imperceptible Adversarial Attacks on 3D Point Clouds (EIA). EIA first identifies edges of the point clouds by detecting locations where point geometric and semantic features exhibit abrupt changes, and then focuses on perturbing these edge points while suppressing perturbation on other points during the attack, thereby reducing distortions. Extensive experiments validate that our method significantly enhances the imperceptibility of the adversarial attack and demonstrates its superiority over existing methods.