MGT-PC: MEMORY-GUIDED TRANSFORMER FOR ROBUST POINT CLOUD CLASSIFICATION

Deep neural network (DNN)-based point cloud classifiers are known to be vulnerable to adversarial examples. This issue exposes the model defects on adversarial robustness and thus threatens security-sensitive systems. We propose a memory-guided transformer for point cloud (MGT-PC) that improves the adversarial robustness of 3D point cloud classification. MGT-PC has three competitive advantages over existing models. First, we enhance MGT-PC with a memory for retaining class-wise information to mitigate robust feature leakage. Second, a transformer-like structure is adopted to refine hidden features with self-attention, which helps defend against global attacks. Third, we propose a contrast loss to sharpen the outputs, which makes the model reduces ambiguity when the model is in a dilemma. Experiments on the ModelNet40 dataset demonstrate that MGT-PC is more robust and less susceptible to adversarial attacks.