Dual-targeted adversarial noise for 3D point cloud classification model

In this paper, we introduce a novel method for generating dual-target adversarial examples in point cloud data, specifically designed to cause different models to misclassify into distinct attacker-specified classes. Although deep-learning models have been used to recognize 3D point clouds, they remain vulnerable to adversarial attacks that exploit the data structure. Our approach addresses this problem by minimizing the loss function using feedback from multiple models, thus ensuring targeted misclassification. This has significant implications, particularly in military contexts, where adversarial examples can manipulate each enemy recognition system into different classes. We validated our method on the ModelNet40 dataset with PointNet and PointNet++ models, demonstrating its effectiveness by visualizing attack success rates, distortions, and point clouds. Experimental results show that the proposed method achieved attack success rates of 99.8% and 84.16% for generating dual-target adversarial examples using the PointNet and PointNet++ models, respectively.