Research map / 3D vision / Gaussian splatting
SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography
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 SecureGS, a Scaffold-GS-based 3DGS steganography framework that hides 3D objects, images, and bits in anchor-point features decoded by private MLPs, with region-aware density optimization to keep hidden geometry out of the visible point cloud.
Problem
3DGS steganography must hide content in a publicly released Gaussian point cloud while preserving rendering fidelity, speed, and security (both file-format and geometric-structure security); prior methods such as GS-Hider expose hidden geometry and render slowly.
Approach
Build on Scaffold-GS anchors: decode original-scene Gaussians with public MLPs and hidden-object Gaussians with private MLPs/offset predictor from shared anchor features, and use decoupled asynchronous gradient accumulation with DBSCAN-located bounding boxes to control anchor growth so hidden anchors stay hidden among original anchors.
Key results
- SecureGS scene-level hiding averages 27.75 dB original-scene PSNR and 32.28 dB hidden PSNR versus GS-Hider's 26.59 and 30.62 dB, at 131.71 FPS versus 48.28 FPS.
- On the Blender bit-hiding comparison, SecureGS reaches 100% bit accuracy versus 92.69% for NeRFProtector and 62.15% for CopyRNeRF.
- Object-level hiding reaches 38.21 dB average hidden PSNR with a 290.54 MB container, while RDO removal yields a 2.21 dB hidden PSNR gain but exposes geometry.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SecureGS instead of GS-Hider | 3D object hiding, scene-level · Original scene and hidden object (Avg.) | PSNR: 26.59 -> 27.75; 30.62 -> 32.28 | Helped |
| use SecureGS instead of GS-Hider | 3D object hiding, scene-level · 3D scenes | FPS: 48.28 -> 131.71 | Helped |
| use SecureGS instead of GS-Hider | 3D object hiding, scene-level · 3D scenes | storage size (MB): 468.63 -> 267.39 | Helped |
| use SecureGS instead of GS-Hider | single-image hiding · Original scene (Avg.) | PSNR: 26.64 -> 27.26 | Helped |
| use SecureGS instead of GS-Hider | single-image hiding · Restored image (Avg.) | PSNR: 42.16 -> 42.99 | Helped |
| use SecureGS instead of NeRFProtector | bit hiding · Blender dataset | bit accuracy: 92.69 -> 100.00 | Helped |
| use SecureGS instead of GS-Hider | 3D scene hiding · 3D scenes | hidden-scene PSNR: 25.18 -> 29.75 | Helped |
| add HDGER (ablation) | SecureGS · 3D object hiding | hidden-scene PSNR: 38.68 -> 38.21 | No clear effect |
| add RDO (ablation) | SecureGS · 3D object hiding | original-scene PSNR: 27.49 -> 27.93 | Helped |
| add RDO (ablation) | SecureGS · 3D object hiding | hidden-scene PSNR: 40.42 -> 38.21 | Hurt |
| add RDO and HDGER (ablation) | SecureGS · 3D object hiding | hidden-scene PSNR: 40.51 -> 38.21 | Hurt |
| lower the gradient threshold globally rather than region-adaptively (ablation) | SecureGS RDO · 3D object hiding | original-scene PSNR: 27.93 -> 28.11 | Helped |
Surprising findings
- Removing the region-aware density optimization raises hidden-object PSNR by 2.21 dB, showing a fidelity-security trade-off.
- Globally lowering the gradient threshold to densify more gives only 0.18 dB original-scene gain while adding about 115 MB.
Limitations
- Storage is slightly larger than Scaffold-GS when hiding 3D objects because RDO promotes anchor splitting.
- Hidden object must partially overlap the original scene's point cloud; hiding in sparse regions is hard to do without leaving traces.
- Bit decoding relies on a dedicated private MLP and does not generalize to arbitrary message extraction like HiDDeN-based decoders.
Experimental setup
Datasets: Mip-NeRF 360 Blender dataset
Benchmarks: Mip-NeRF 360 Blender dataset
Metrics: PSNR SSIM LPIPS FPS storage size (MB) bit accuracy
Baselines: GS-Hider 3DGS+StegaNeRF CopyRNeRF NeRFProtector Scaffold-GS
Builds on: Scaffold-GS 3D Gaussian Splatting GS-Hider StegaNeRF DBSCAN
Closest papers by meaning
- GS-Hider: Hiding Messages into 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.94)
- GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- Geometry Cloak: Preventing TGS-based 3D Reconstruction from Copyrighted Images (NeurIPS 2024 · similarity 0.90)
- GuardSplat: Efficient and Robust Watermarking for 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- 3D-GSW: 3D Gaussian Splatting for Robust Watermarking (CVPR 2025 · similarity 0.88)
- Poison-splat: Computation Cost Attack on 3D Gaussian Splatting (ICLR 2025 · similarity 0.87)
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.