Segments-aware Universal Adversarial Perturbations Purification on 3D Point Cloud Classifiers

Yang GaoXianrui ChangHaoran LiJian Xu^*

• Northeastern University, Shenyang, China

3D point cloud classifiers, while powerful for representing real-world objects and environments, are vulnerable to adversarial perturbations, particularly Universal Adversarial Perturbations (UAPs) which pose a significant threat due to their input-agnostic nature. Existing purification methods often operate independently of the target classifier and treat perturbations as isolated points, failing to consider the often coherent, structural nature of UAPs in 3D point clouds (e.g., outlier-like shapes with continuous curvature). This oversight limits their effectiveness, especially since distinguishing between genuine geometry and structured adversarial patterns is challenging. To overcome these limitations, we propose a novel purification framework that leverages model interpretability to identify and remove adversarial regions. Our approach uniquely identifies influential regions within adversarial samples that maximally impact the classifier's predictions. Recognizing that UAPs often manifest as structured segments, we employ graph wavelet transforms to isolate suspicious curvature segments. These segments are then subjected to a transplantation test: they are considered adversarial if transferring them to clean samples consistently induces misclassification. Adversarial regions identified through this process are subsequently removed to sanitize the point cloud. This model-guided, structure-aware approach treats UAPs holistically rather than as individual points. We demonstrate the effectiveness of our method through extensive experiments on two public datasets and four different 3D point cloud classifiers, showcasing remarkable improvements in robustness against UAPs.