Generating Transferable Traffic Object Adversarial 3D Point Clouds via Momentum-based Decompose Perturbation

Abstract. With the rapid development of mobile mapping technology, 3D point cloud data is widely used in the field of intelligent driving. In intelligent driving systems, the recognition ability of point cloud objects is crucial for achieving safe driving. However, existing deep neural networks are prone to making incorrect judgments when subjected to adversarial attacks, which may lead to serious consequences. Most of the existing point cloud perturbation methods are based on white box attacks and cannot successfully attack models with unknown parameters, which is still different from real usage scenarios. In this paper, we focus on studying the transferability of point cloud perturbation, that is, successful attacks on a model can also be transferred to models that have not participated in generating perturbations, making them make incorrect judgments. We propose a new method for generating adversarial point clouds, named MBDP, which decomposes the adversarial point cloud into two sub-perturbations using the decomposition perturbation method. The momentum iterative fast sign algorithm is used to optimize both the sub-perturbation and the main-perturbation simultaneously, generating adversarial samples that are far from the decision boundary and more transferable. Experimental results show that both on real and synthetic 3D datasets, our proposed MBDP achieve the highest attack success rate and transferability score.