Enhancing robustness of point cloud semantic segmentation against adversarial attacks using silhouette coefficient regularized neural ODEs

• Main Author: Hong, Jianxiong
• Other Authors: Tay Wee Peng
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Computer and Information Science; Point cloud semantic segmentation; Adversarial attacks; Adversarial defense; Neural ODEs; Silhouette coefficient regularization; SODEF
• Online Access: https://hdl.handle.net/10356/182637
• Institution: Nanyang Technological University

This dissertation aims to enhance the robustness of point cloud semantic seg mentation against adversarial attacks by proposing an improved defense mech anism. The primary focus is on developing a novel model, termed Silhou ette Coefficient Regularization Augmented Stable Neural ODE with Lyapunov Stable Equilibrium Points (SA-SODEF). The traditional Stable Neural ODE with Lyapunov-Stable Equilibrium Points (SODEF) method, while effective in defend ing against adversarial attacks in image classification tasks, has limitations when applied to unbalanced point cloud datasets. To address these limitations, SA-SODEF incorporates the silhouette coefficient as a regularization term and introduces a trainable fully connected layer in the SODEF block. This combination enables SA-SODEF to better adapt to the dis tribution of feature points produced by the feature extractor and to handle un balanced classes more effectively. Through extensive experiments on two urban scene datasets, the Semantic Ur ban Meshes (SUM) and the Hessigheim3D (H3D), the results demonstrate that SA-SODEF consistently outperforms SODEF in terms of robustness against Fast Gradient Sign Method (FGSM) attacks. The introduction of the silhouette coef ficient regularization term and the trainable fully connected layer optimizes the arrangement of equilibrium points, ensuring that larger classes occupy propor tionally larger areas in the feature space. This contributes to a more robust and balanced feature space representation, making SA-SODEF more resilient to adversarial perturbations. In summary, this dissertation presents SA-SODEF as a significant improvement over SODEF in enhancing the robustness of point cloud semantic segmentation against adversarial attacks, particularly on unbalanced datasets. Keywords: Point Cloud Semantic Segmentation, Adversarial Attacks, Adversar ial Defense, Neural ODEs, Silhouette Coefficient Regularization, SODEF