Shape Prior Guided Attack: Sparser Perturbations on 3D Point Clouds

Deep neural networks are extremely vulnerable to malicious input data. As 3D data is increasingly used in vision tasks such as robots, autonomous driving and drones, the internal robustness of classification models for point cloud has received widespread attention. In this paper, we propose a novel method named SPGA (Shape Prior Guided Attack) generate adversarial examples. We use shape prior information make perturbations sparser thus achieve imperceptible attacks. particular, Spatially Logical Block (SLB) apply points through sliding oriented bounding box. Moreover, design an algorithm called FOFA type task, which further refines attack process breaking down complicated problems into sub-problems. Compared with methods global perturbation, our consumes significantly fewer computations, making it more efficient. Most importantly all, can examples higher success rate (even defensive situation), less perturbation budget stronger transferability.