Beyond Defenses: Manifold-Aligned Regularization for Intrinsic 3D Point Cloud Robustness
Researchers have developed a new method called Manifold-Aligned Point Recognition (MAPR) to improve the robustness of 3D point cloud networks against adversarial attacks. MAPR addresses the issue of latent geometry misalignment by regularizing the network's feature space to be invariant to intrinsic, geometry-preserving perturbations. This approach does not require adversarial training or additional data, yet it significantly enhances robustness on benchmark datasets like ModelNet40 and ScanObjectNN, showing average gains of over 20 and 8 percentage points respectively. AI
IMPACT Enhances the security and reliability of 3D AI models against malicious manipulation.