3D-ANC : Adaptive Neural Collapse for Robust 3D Point Cloud Recognition
Yuanmin Huang, Wenxuan Li, Mi Zhang, Xiaohan Zhang, Xiaoyu You, Min Yang · 2025
Read 3D-ANC : Adaptive Neural Collapse for Robust 3D Point Cloud Recognition by Yuanmin Huang, Wenxuan Li, Mi Zhang, Xiaohan Zhang, Xiaoyu You, Min Yang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-07040.txt.
Description
Yuanmin Huang and five co-authors defend 3D point cloud recognition against adversarial perturbations, small crafted changes that fool classifiers and threaten real deployments. They argue current defenses fail mainly because their feature spaces are entangled, leaving room for attacks. Their method, 3D-ANC, instead uses neural collapse, the tendency of last-layer features to settle into a simplex arrangement of maximally separated class prototypes.
Class imbalance and geometric similarity between object categories complicate that plan for point clouds. So the method adds an aligned classification module plus adaptive training with representation-balanced learning and a dynamic feature direction loss. The payoff is large: on ModelNet40, one backbone's accuracy under attack rose from about twenty-seven percent to about eighty-one percent, far beyond leading baselines, and gains held across model structures.
Catalog record
| CREATOR | Yuanmin Huang, Wenxuan Li, Mi Zhang, Xiaohan Zhang, Xiaoyu You, Min Yang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-07040