Research map / 3D vision / Gaussian splatting
3D-GSW: 3D Gaussian Splatting for Robust Watermarking
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 3D-GSW, a robust watermarking method that fine-tunes 3D Gaussian Splatting models to embed a binary message in rendered images, using frequency-guided densification, a gradient mask, and a wavelet-subband loss.
Problem
Watermarking 3D Gaussian Splatting models and their rendered images to protect copyright against image distortions and model-level attacks, while keeping rendering quality high and rendering real-time.
Approach
Before fine-tuning, remove 3D Gaussians with minimal contribution to rendering quality and split Gaussians in high-frequency patches (found via patch-wise DFT), then fine-tune with a gradient mask derived from pre-trained parameters and a wavelet high-frequency subband loss, embedding the message via a frozen HiDDeN decoder on the DWT LL2 subband.
Key results
- At 32 bits, 3D-GSW reaches 97.37% bit accuracy and 35.08 PSNR, versus 93.42% and 30.49 PSNR for WateRF+3D-GS.
- Under model distortion (removing 20% of Gaussians), 3D-GSW keeps 96.87% bit accuracy vs. 80.58% for WateRF+3D-GS.
- On Mip-NeRF 360 at 64 bits, fine-tuning takes 21m 03s vs. 6h 47m for WateRF+3D-GS, and FPS rises from 56.65 to 72.68.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FGD to 3D-GSW (ablation) | 3D-GSW · Blender, LLFF, Mip-NeRF 360 (average), 32 bits | Bit accuracy: 94.60 -> 97.37 | Helped |
| Add FGD to 3D-GSW (ablation) | 3D-GSW · Blender, LLFF, Mip-NeRF 360 (average), 64 bits | PSNR: 28.22 -> 32.47 | Helped |
| Add the gradient mask (ablation) | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | PSNR: 29.96 -> 35.08 | Helped |
| Add the wavelet-subband loss (ablation) | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | PSNR: 33.26 -> 35.08 | Helped |
| Use only high-frequency wavelet subbands (LH, HL, HH) rather than all subbands (ablation) | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | Bit accuracy: 96.01 -> 97.37 | Helped |
| Use only high-frequency wavelet subbands (LH, HL, HH) rather than all subbands (ablation) | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | PSNR: 34.93 -> 35.08 | Helped |
| Increase message length from 32 bits to 64 bits | 3D-GSW · Blender, LLFF, Mip-NeRF 360 (average) | Bit accuracy: 97.37 -> 90.45 | Hurt |
| Apply image distortions during extraction | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits; Gaussian noise (σ = 0.1) | Bit accuracy: 97.37 -> 83.84 | Hurt |
| Apply rotation during extraction | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits; rotation (±π/6) | Bit accuracy: 97.37 -> 87.94 | Hurt |
| Add Gaussian noise to 3D-GS parameters (σ = 0.1) | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | Bit accuracy: 97.37 -> 89.11 | Hurt |
| Remove 3D Gaussians (20%) from the model | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | Bit accuracy: 97.37 -> 96.87 | Hurt |
| Clone 3D Gaussians (20%) in the model | 3D-GSW · Blender, LLFF, Mip-NeRF 360, 32 bits | Bit accuracy: 97.37 -> 95.99 | Hurt |
Surprising findings
- Removing about 28% of 3D Gaussians slightly increased bit accuracy while keeping rendering quality, and removing more than 50% degraded bit accuracy and LPIPS.
Limitations
- Requires a pre-trained message decoder for each bit length, so decoder pre-training must be done first.
- Robustness and capacity trade off: bit accuracy and rendering quality decline as message length grows from 32 to 64 bits.
Experimental setup
Datasets: Blender LLFF Mip-NeRF 360
Benchmarks: Blender LLFF Mip-NeRF 360
Metrics: bit accuracy PSNR SSIM LPIPS FPS storage fine-tune time
Baselines: StegaNeRF+3D-GS WateRF+3D-GS 3D-GSW without FGD
Builds on: 3D Gaussian Splatting WateRF StegaNeRF HiDDeN error-based densification
Closest papers by meaning
- GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.92)
- GuardSplat: Efficient and Robust Watermarking for 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- GS-Hider: Hiding Messages into 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography (ICLR 2025 · similarity 0.88)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
- Geometry Cloak: Preventing TGS-based 3D Reconstruction from Copyrighted Images (NeurIPS 2024 · 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.