Research map / Point clouds / Point cloud classification
Unlearnable 3D Point Clouds: Class-wise Transformation Is All You Need
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UMT, a class-wise multi-transformation unlearnable scheme for 3D point clouds with a matrix-inverse data restoration step so authorized users can still train, validated on six datasets and 16 models.
Problem
Unauthorized training on 3D point cloud data leaks sensitive information, and existing unlearnable-example methods from 2D either do not transfer to point clouds or also block authorized users from learning from protected data.
Approach
Assign class-wise combinations of invertible 3D transformations (rotation, scaling, shear, twisting) to point cloud samples so models learn a transformation-to-label shortcut, and let authorized users undo the transformation with class-wise inverse matrices sent as a lightweight message from the protector.
Key results
- On ModelNet40 UMT with RS reduces average test accuracy across 16 models to about 13-14% versus about 86% for clean training (table of UMT variants).
- On S3DIS semantic segmentation, UMT lowers PointNet++ eval accuracy from 74.76% to 15.89% and Point Transformer v3 mIoU from 40.57% to 18.32%.
- Adaptive attack with random RS augmentation keeps accuracy about 28.67% below clean baseline on the reported setting, showing partial robustness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply class-wise multi-transformations to training data | Point cloud classifiers trained with UMT · ModelNet40 | test accuracy (%): 85.89 -> 13.19 | Helped |
| Apply class-wise multi-transformations to training data | Point cloud classifiers trained with UMT · ModelNet10 | test accuracy (%): 89.61 -> 23.19 | Helped |
| Apply class-wise multi-transformations to training data | Segmentation models trained with UMT · S3DIS | mIoU (%): 43.63 -> 24.02 | Helped |
| Add scaling to rotation-based UMT (ablation) | UMT · ModelNet40 | test accuracy (%): 33.84 -> 15.04 | Helped |
| Add rotation to scaling-based UMT (ablation) | UMT · ModelNet40 | test accuracy (%): 30.22 -> 15.04 | Helped |
| Add scaling to rotation-based UMT (ablation) | UMT · ModelNet10 | test accuracy (%): 42.66 -> 25.78 | Helped |
| Add rotation to scaling-based UMT (ablation) | UMT · ModelNet10 | test accuracy (%): 47.43 -> 25.78 | Helped |
| Apply random rotation and scaling as an adaptive preprocessing attack | UMT · UMT-ModelNet40 | test accuracy (%): 14.73 -> 56.19 | Hurt |
| Apply class-wise rather than sample-wise or dataset-wise parameter assignment | PointNet trained on ModelNet10 · ModelNet10 | test accuracy (%): 89.32 -> 45.14 | Helped |
Surprising findings
- Rigid transformations such as rotation and scaling are easily defeated by rotation- or scale-invariant networks such as RIConv, LGR-Net, and 3DGCN.
- Sample-wise and dataset-wise transformations do not reduce accuracy and can act as augmentation, so only class-wise settings work as unlearnable schemes.
Limitations
- Rigid transformations can be defeated by invariant networks, and the paper suggests adding non-rigid transforms for robustness.
- Robustness to adaptive attacks is only partial, with accuracy still notably below clean baseline.
- Theoretical analysis is restricted to a binary GMM setting with the RS transformation.
Experimental setup
Datasets: ModelNet40 ModelNet10 ShapeNetPart KITTI ScanObjectNN S3DIS
Benchmarks: ModelNet10 ModelNet40 ShapeNetPart KITTI ScanObjectNN S3DIS
Metrics: test accuracy eval accuracy mIoU
Baselines: Clean baseline SOR SRS Random rotation Random scaling Random jitter
Builds on: unlearnable examples Gaussian Mixture Model Bayes classifier analysis
Closest papers by meaning
- ConcreTizer: Model Inversion Attack via Occupancy Classification and Dispersion Control for 3D Point Cloud Restoration (ICLR 2025 · similarity 0.86)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.84)
- Toward Availability Attacks in 3D Point Clouds (ICML 2024 · similarity 0.83)
- An Explicit Frame Construction for Normalizing 3D Point Clouds (ICML 2024 · similarity 0.83)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · similarity 0.83)
- Improving Robustness of 3D Point Cloud Recognition from a Fourier Perspective (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.