Invisible Adversarial Attack Against 3D Point Cloud Classifier Based on Curvature Constrain

With the widespread adoption of 3D point cloud technology in autonomous driving, robotic navigation, and similar fields, the security of point cloud data has become increasingly critical. Traditional gradient-based adversarial attacks on point clouds, such as the Iterative Fast Gradient Sign Method, often compromise imperceptibility by producing adversarial samples with noticeable geometric distortions while pursuing high attack success rates, making them easily detectable by humans or systems. This paper introduces the Curvature-Constrained Iterative Fast Gradient Sign Method, which enhances the i-FGSM framework by incorporating a multi-objective loss function to jointly optimize adversarial loss and local curvature constraints, generating samples that balance high attack efficacy with strong imperceptibility. The results demonstrate that CC-iFGSM significantly reduces Mean Curvature Variation and Maximum Curvature Variation, with the generated adversarial samples outperforming baseline methods in both imperceptibility and visual quality.