Comrade-Secure Adversarial Noise for 3D Point Cloud Classification Model

Abstract

Deep neural networks (DNNs) are effective across many domains, including text, audio, and image. Recently, DNNs have been used in autonomous driving, robotics, and even drones owing to the increasing utilization of 3D data. However, 3D data point clouds are vulnerable to adversarial examples, much like any other form of data. An adversarial example slightly alters the original sample or adds a small amount of noise, making it appear normal to humans, which results in its misclassification by the models. In this study, we propose a method that can be used to generate a “comrade-secure” adversarial point cloud example. In the proposed method, we subtly adjust the positions of certain points in the point cloud to create an adversarial example. This alteration causes the enemy model to misclassify, while the friendly model remains accurate. We use the ModelNet40 dataset for experimental evaluation and utilize PointNet++ and PointNet, which are representative models to classify 3D point clouds, as friendly and enemy models, respectively. In the experiments, the adversarial point cloud examples generated by the proposed method showed that the friendly model achieved an accuracy of 97.65%, and the enemy model was misclassified with an attack success rate of 99.55%.