Saliency-Aware Generation of Adversarial Point Clouds
Lianguang Liu, Zeping Ye, Junting Lv, Zhigang Zeng · 2023
Deep learning for 3D point clouds has achieved excellent performance in some safety-critical applications, such as self-driving. The robustness of deep 3D models when facing adversarial attacks has attracted more and more attention. Existing 3D attack methods design a variety of perturbation constraints to improve the imperceptibility of the adversarial point clouds. However, these constraints generally lack consideration for the saliency degree of the points. We are the first to enhance the imperceptibility of adversarial samples by combining saliency maps of point clouds with the perturbation metrics. Moreover, we introduce Earth Mover’s Distance (EMD) to better measure the overall perturbation size of the point clouds. Finally, a novel saliency attack is proposed to combine the saliency-aware perturbation constraints and EMD. Experimental results show that our method can efficiently generate imperceptible 3D adversarial samples with tinny perturbations.