Anchor Field Consistency for Imperceptible Adversarial Attacks on 3D Point Clouds

Adversarial attacks on point clouds are crucial for assessing and improving the robustness of 3D deep models. Prior attacks mostly enforce imperceptibility using set-level distances, e.g., the Chamfer and Hausdorff distances, yet visible artifacts remain. We attribute this to nearest-neighbor correspondence sensitivity when comparing clean and adversarial shapes. To address this, we propose Anchor Field Consistency (AFC), which evaluates the clean and adversarial shapes at the same anchors. Specifically, at each anchor we approximate a local surface projection, construct a compact distance–direction representation, and enforce consistency across shapes during optimization. Extensive experiments show that attacks augmented with AFC generate more surface-coherent adversarial point clouds and outperform state-of-the-art attack methods.