3D adversarial attacks beyond point cloud

Recently, 3D deep learning models have been shown to be susceptible adversarial attacks like their 2D counterparts. Most of the state-of-the-art (SOTA) perform perturbation point clouds. To reproduce these in physical scenario, a generated cloud needs reconstructed mesh, which leads significant drop its effect. In this paper, we propose strong attack named Mesh Attack address problem by directly performing on mesh object. order take advantage most effective gradient-based attack, differentiable sample module that back-propagate gradient is introduced. further ensure examples without outlier and printable, three losses are adopted. Extensive experiments demonstrate proposed scheme outperforms SOTA margin. We also achieved performance under various defenses. Our code will available at: https://github.com/cuge1995/Mesh-Attack.