Research map / Point clouds / Point cloud classification
Transferable 3D Adversarial Shape Completion using Diffusion Models
Xuelong Dai, Bin Xiao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ensemble adversarial guidance and model uncertainty to 3DAdvDiff | 3DAdvDiff · ShapeNet Chair, black-box transfer | average ASR (%): 19.0 -> 85.2 | Improved |
| add model uncertainty via random down-sampling (ablation) | 3DAdvDiff · ShapeNet Chair | average black-box ASR (%): 19.0 -> 75.5 | Improved |
| add ensemble adversarial guidance (ablation) | 3DAdvDiff · ShapeNet Chair | average black-box ASR (%): 19.0 -> 65.2 | Improved |
| increase number of random samples M (ablation) | 3DAdvDiff model uncertainty · ShapeNet | transferability and running time | Mixed |
| replace perturbation-based attack with diffusion shape completion (white-box 3DAdvDiff) | PF-Attack / PGD baselines · ShapeNet All, black-box transfer | average black-box ASR (%): 6.2 -> 36.6 | Improved |
| integrate 3DAdvDiff boosting methods into PGD without diffusion | PGD · ShapeNet whole test set | average black-box ASR (%): 9.2 -> 25.0 | Improved |
| integrate 3DAdvDiff boosting methods into SI-Adv | SI-Adv · ShapeNet whole test set | PointNet ASR (white-box): 97.6 -> 70.5 | Hurt |
| use ensemble guidance with critical-point selection vs plain 3DAdvDiff | 3DAdvDiff · IF-Defense (PointNet) | ASR (%): 30.0 -> 22.5 | Hurt |
| use ensemble version against defenses SRS, SOR, DUP-Net, Hybrid Training | 3DAdvDiff · Hybrid Training defense | ASR (%): 9.4 -> 96.1 | Improved |
| add generation quality augmentation (l0 limit on critical points) (ablation) | 3DAdvDiff_ens · ShapeNet | generation quality and ASR | Improved |
| diffusion-based generation instead of perturbation attack | 3DAdvDiff_ens · ShapeNet | running time (s): 1.1 -> 60.8 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdvPC — AdvPC: Transferable Adversarial Perturbations on 3D Point Clouds (ECCV 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards a 3D Transfer-based Black-box Attack via Critical Feature Guidance (ICCV 2025 · similarity 0.93)
- Generating Transferable 3D Adversarial Point Cloud via Random Perturbation Factorization (AAAI 2023 · similarity 0.92)
- Hide in Thicket: Generating Imperceptible and Rational Adversarial Perturbations on 3D Point Clouds (CVPR 2024 · similarity 0.91)
- Curvature-Invariant Adversarial Attacks for 3D Point Clouds (AAAI 2024 · similarity 0.90)
- Explicitly Perceiving and Preserving the Local Geometric Structures for 3D Point Cloud Attack (AAAI 2024 · similarity 0.89)
- Towards Transferable Targeted 3D Adversarial Attack in the Physical World (CVPR 2024 · similarity 0.89)
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.
