中文超越防御：面向内在三维点云鲁棒性的流形对齐正则化

尽管点云鲁棒性研究已取得进展，现有方法多依赖数据增强或防御机制，却忽视了对抗脆弱性的几何本质。研究发现，3D网络的对抗脆弱性源于模型学习的潜在几何与底层表面内在几何之间的流形错位：沿输入流形的微小保几何扰动会在特征空间引发不成比例的扭曲。这一发现揭示了对抗攻击的几何根源，为设计更鲁棒的3D深度学习模型提供了新思路。

arXiv:2605.07590v2 Announce Type: replace Abstract: Despite extensive progress in point cloud robustness, existing methods primarily rely on augmentation strategies or defense mechanisms while overlooking the geometric nature of adversarial fragility. We hypothesize that adversarial vulnerability in 3D networks arises from a manifold misalignment between the latent geometry learned by the model and the intrinsic geometry of the underlying surface. Small, geometry-preserving perturbations along the input manifold often induce disproportionate distortions in feature space, potentially leading to