Curvature-Guided Dynamic Search for Imperceptible Adversarial Attacks on Point Clouds

Deep learning models for 3D point cloud analysis have demonstrated significant vulnerability to adversarial attacks, raising profound security concerns in consumer electronics applications, including AR/VR devices, domestic robotics, and autonomous driving systems. While existing 3D attack methods primarily enforce implicit constraints on point perturbations via geometry-aware loss functions, they frequently suffer from noticeable outliers and geometric distortion. Although recent efforts have incorporated surface priors for explicit constraints to mitigate such artifacts, two fundamental limitations remain: 1) static reliance on local geometric information fails to capture the intricate topological structures of point cloud surfaces, preventing perturbations from seamlessly adapting to complex geometries; and 2) uniform perturbation across the entire point cloud neglects the heterogeneity in gradient sensitivity among points, compromising overall stealthiness. To overcome these bottlenecks, we propose the Curvature-Guided Dynamic Search (CDS) framework for crafting visually imperceptible adversarial point clouds. CDS introduces a dynamic curvature-guided path mechanism that iteratively expands the search neighborhood along the minimum principal curvature direction, enabling perturbations to naturally conform to complex surface topologies. Concurrently, a gradient sensitivity field concentrates perturbation energy on the most attack-critical points while preserving global shape integrity. Extensive experiments demonstrate that CDS substantially enhances geometric stealthiness while maintaining a formidable attack success rate, consistently outperforming state-of-the-art methods.