Abstract:

Deep learning-based 3D point cloud understanding has received increasing attention in various applications such as autonomous driving, robotics, surveillance, etc., and the study of adversarial attacks on point cloud deep learning models helps to evaluate and improve their adversarial robustness. However, most of the existing attack methods are aimed at white-box attacks, generating adversarial samples that have very low success rate and are easily defensible against transferable attacks on black-box models with unknown model parameters. These methods only consider optimization in the geometric space to mislead specific classifiers and fail to essentially change the deep intrinsic semantic structure of point cloud data, resulting in their limited ability to transferable attacks under different classifiers. To address these issues, the proposed algorithm leveraged the rich semantic comprehension capability of large multimodal models to incorporate the semantic information of the point clouds into the attack, thereby ensuring that the adversarial samples diverged significantly from the original semantic attributes to a remarkable extent to enhance transferability. In addition, considering that the current adversarial samples with high attack transferability often exhibited insufficient imperceptibility, the algorithm integrated the above semantic adversarial attack into the spectral domain space, achieving a delicate balance between transferability and imperceptibility. Extensive evaluations demonstrated the 3D CLIP-based semantic offset attack (3DCLAT) can significantly improve the transferability of the adversarial samples and is more robust to defense methods.

Key words: CLIP, point cloud, adversarial attack, attack transferability, spectral domains