Dataset for "Toward Leveraging Intrinsic Point Cloud Features in 3D Adversarial Attacks" paper

This dataset accompanies the paper “Toward Leveraging Intrinsic Point Cloud Features in 3D Adversarial Attacks” and provides the data used in the experimental analysis. This dataset provides point-wise geometric features and saliency information extracted from the ModelNet40 test set, used for analyzing adversarial point cloud attacks. The dataset includes two NumPy (.npy) files containing handcrafted geometric features and model-based saliency scores for each 3D point. modelnet40_test_xyz_f1to14.npy This file contains point-wise geometric feature data extracted from the ModelNet40 test set, comprising 2468 point clouds. For each 3D point: • Columns 1–3 store the normalized Cartesian coordinates (x, y, z). • Columns 4–17 correspond to the fourteen handcrafted geometric features f₁ to f₁₄ defined in the paper. Among these features, f14 (distance from the centroid) is used as the the proposed adversarial drop attack according to paper. Higher values indicate stronger geometric feature presence for the corresponding point. modelnet40_test_saliency_4models.npy This file contains point-wise saliency scores computed on the ModelNet40 test set for four different DNN architectures. For each 3D point: • Columns 1–3 store the normalized Cartesian coordinates (x, y, z). • Columns 4–7 store saliency scores computed using: • PointNet • PointNet++ • DGCNN • PointConv , respectively. Higher values indicate higher adversarial relevance of the corresponding point