Research map / 3D vision / Gaussian splatting
Geometry Cloak: Preventing TGS-based 3D Reconstruction from Copyrighted Images
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 "geometry cloaks", imperceptible adversarial perturbations added to input images that make Triplane Gaussian Splatting (TGS) produce a customized watermark pattern in a chosen view, enabling copyright claims against unauthorized single-view 3D reconstruction.
Problem
Single-view 3D reconstruction methods like TGS can generate 3D models from copyrighted images without authorization, and existing adversarial or watermarking approaches either only degrade output or cannot transfer watermarks into reconstructed 3D models.
Approach
Target the explicit point-cloud geometry features of TGS (rather than image features) with a view-specific projected gradient descent that minimizes Chamfer distance between the projected reconstructed point cloud and a predefined or customized watermark pattern, so the reconstructed 3D model reveals the watermark when viewed from a chosen direction.
Key results
- On Omni3D at epsilon=2, geometry cloak reduces PSNR to 12.47 and raises CD to 337.0, versus 23.35 PSNR and CD 4.913 for Gaussian noise.
- Attacking image features (Adv. image) barely changes reconstructions (CD 13.86 at epsilon=2 on Omni3D), supporting the claim that TGS image features are robust while geometry is perturbation-prone.
- The method extends to LGM, with CD rising to 132.2-185.0 under the geometry cloak versus about 1.4-4.9 for Gaussian noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a target geometry pattern instead of attacking point-cloud features without a target (ablation) | Geometry cloak on TGS · Omni3D, ε=4 | CD: 260.1 -> 344.9 | Helped |
| Add a target geometry pattern instead of attacking point-cloud features without a target (ablation) | Geometry cloak on TGS · GSO, ε=4 | CD: 284.8 -> 356.5 | Helped |
| Attack explicit point-cloud geometry rather than implicit image features (ablation) | TGS · Omni3D, ε=4 | CD: 14.84 -> 260.1 | Helped |
| Attack explicit point-cloud geometry rather than implicit image features (ablation) | TGS · GSO, ε=4 | CD: 13.60 -> 284.8 | Helped |
| Apply geometry cloak instead of Gaussian noise | TGS · Omni3D, ε=4 | CD: 6.424 -> 344.9 | Helped |
| Apply geometry cloak instead of Gaussian noise | TGS · GSO, ε=4 | CD: 7.940 -> 356.5 | Helped |
| Apply geometry cloak instead of Gaussian noise | LGM · Anya | CD: 4.921 -> 132.2 | Helped |
| Apply geometry cloak instead of Gaussian noise | LGM · Bird | CD: 3.506 -> 185.0 | Helped |
| Apply geometry cloak instead of Gaussian noise | LGM · Catstatue | CD: 1.430 -> 125.3 | Helped |
| Combine tri-plane-feature perturbation with geometry-feature perturbation (ablation) | Geometry cloak on TGS · 3D reconstruction | attack performance | No clear effect |
Surprising findings
- Perturbing image features of TGS barely affects its 3D reconstructions, while small geometry-level perturbations drastically change them.
- Combining tri-plane perturbation with geometry cloak does not improve attack performance, since the tri-plane is robust.
Limitations
- Limitations section is only a generic discussion of responsible use; no concrete technical limitations are stated.
- Robustness to image compression and common operations is only partially reported, and the watermark is only visible from specific views.
Experimental setup
Datasets: Google Scanned Objects (GSO) OmniObject3D (Omni3D)
Benchmarks: Google Scanned Objects (GSO) OmniObject3D (Omni3D)
Metrics: PSNR SSIM LPIPS Chamfer Distance
Baselines: Gaussian noise Adv. image Geometry cloak w/o target
Builds on: Triplane Gaussian Splatting projected gradient descent adversarial attacks on NeRF
Closest papers by meaning
- GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography (ICLR 2025 · similarity 0.90)
- GS-Hider: Hiding Messages into 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- GuardSplat: Efficient and Robust Watermarking for 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- 3D-GSW: 3D Gaussian Splatting for Robust Watermarking (CVPR 2025 · similarity 0.87)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.84)
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.