Adversarial Attack Against 3D Shapes Utilizing Their Common Points

3D shape, whose representation contains 3D point cloud and 3D mesh, has played an important role in many security-sensitive domains with the application of 3D Deep Neural Network (DNN). However, adversarial attack aims to threaten DNNs' security by misleading them into wrong prediction. Existing adversarial attacks are primarily designed for point cloud with few studies on 3D mesh adversarial attack. Such disproportion leads to the limitation of the practicality of 3D adversarial attack. Therefore, we propose a 3D adversarial attack, named AdvSCP, which perturbs on the common points to support both 3D point cloud and 3D mesh. In detail, we first design the generation of adversarial perturbation according to the direction of back-propagated gradient. Then we apply the perturbation on the overlapping coordinate field for two reasons. To enhance the effectiveness and stealthiness of our method, iterative generation on adversarial result and append 3D distance measurement as constraint in loss function are considered. Experiments on AdvSCP achieve strong attack efficiency while maintaining imperceptible stealthiness.