Adversarial Geometric Transformations of Point Clouds for Physical Attack

Towards adversarial physical attack in real world, we argue that the main challenge lies in discounting adversarial effects by changes of point density along object surface. Most of existing point-wise perturbation based attackers concern on suppressing geometric irregularities, but it remains challenging to produce adversarial shape with geometric smoothness. Adversarial attack via the isometry transformation can alleviate irregular geometries but suffer from its rotation-sensitive nature, so its impractical assumption of category-level pre-alignment on benign object point clouds cannot be relaxed. In light of this, we explore non-rigid geometric transformations for geometry-aware adversaries with a flexible density-aware transformation on the whole point sets, which can thus impose constraints of global and local surface properties when adversarially deforming points. Experiment results on publicly benchmarking ModelNet40 and ScanObjectNN datasets verify the effectiveness of our transformation-based generation algorithms for adversarial shape and physical attack against both rotation sensitive and agnostic point classifiers, significantly outperforming existing adversarial point attackers under diverse recent defenses and the state-of-the-art physical attack methods.