Research map / 3D vision / Gaussian splatting
GS-Hider: Hiding Messages into 3D Gaussian Splatting
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 GS-Hider, a steganography framework that hides 3D scenes or images inside a 3D Gaussian Splatting point cloud by replacing SH colors with a coupled high-dimensional feature attribute decoded by two parallel convolutional decoders.
Problem
Protecting copyright and enabling covert messages in explicit 3DGS point cloud files, which are public and whose attributes are transparent, so naive attribute or decoder-based embedding harms security and fidelity.
Approach
Replace each Gaussian's 48-dim spherical harmonic color with a learnable M=16-dim coupled secured feature that is rasterized into a feature map; a public scene decoder recovers the original RGB view while a private message decoder recovers the hidden scene, so the embedded content is entangled and indistinguishable within the point cloud.
Key results
- GS-Hider averages 25.817 dB PSNR on original scenes and 25.179 dB on hidden scenes, versus 27.205 dB for 3DGS and 24.689/23.232 dB for 3DGS+SH, at about half the storage (411 MB vs 796 MB).
- Single-image hiding reaches 42.532 dB PSNR for the hidden image with 26.440 dB original-scene PSNR.
- Rendering runs at about 45 fps, above the 30 fps real-time threshold, per the paper's measurement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use coupled secured features with parallel scene and message decoders instead of 3DGS+SH | GS-Hider · Mip-NeRF360 (9-scene average) | PSNR_S: 24.689 -> 25.817 | Helped |
| Use coupled secured features with parallel scene and message decoders instead of 3DGS+SH | GS-Hider · Mip-NeRF360 (9-scene average) | PSNR_M: 23.232 -> 25.179 | Helped |
| Set λ to 0.25 rather than 0.5 (ablation) | GS-Hider · Mip-NeRF360 single 3D-scene hiding | PSNR_S: 25.817 -> 26.156 | Helped |
| Set λ to 0.25 rather than 0.5 (ablation) | GS-Hider · Mip-NeRF360 single 3D-scene hiding | PSNR_M: 25.179 -> 19.837 | Hurt |
| Set λ to 1.0 rather than 0.5 (ablation) | GS-Hider · Mip-NeRF360 single 3D-scene hiding | PSNR_M: 25.179 -> 28.802 | Helped |
| Set λ to 1.0 rather than 0.5 (ablation) | GS-Hider · Mip-NeRF360 single 3D-scene hiding | PSNR_S: 25.817 -> 24.932 | Hurt |
| Set feature dimension M to 16 rather than 8 (ablation) | GS-Hider · Mip-NeRF360 single 3D-scene hiding | PSNR_M: 25.102 -> 25.179 | Helped |
| Set feature dimension M to 16 rather than 32 (ablation) | GS-Hider · Mip-NeRF360 single 3D-scene hiding | PSNR_M: 24.547 -> 25.179 | Helped |
| Use 5 rather than 3 Conv+ReLU layers (ablation) | GS-Hider decoders · Mip-NeRF360 single 3D-scene hiding | PSNR_M: 24.306 -> 25.179 | Helped |
| Use 5 rather than 7 Conv+ReLU layers (ablation) | GS-Hider decoders · Mip-NeRF360 single 3D-scene hiding | PSNR_M: 25.103 -> 25.179 | Helped |
Surprising findings
- Random pruning up to 25% degrades original-scene PSNR by only about 2 dB, and sequential opacity pruning has minimal effect.
Limitations
- Rendering quality is marginally below original 3DGS because view dependency is not modeled.
- Rendering speed is reduced by high-dimensional feature rasterization and the multi-layer CNN decoder.
- Hidden-scene PSNR is lower than the original-scene PSNR for many scenes, and some hidden scenes have low SSIM (e.g., Treehill 0.464).
Experimental setup
Datasets: Mip-NeRF360
Benchmarks: Mip-NeRF360
Metrics: PSNR SSIM LPIPS ROC under StegExpose rendering time (fps)
Baselines: 3DGS 3DGS+SH 3DGS+Decoder StegExpose
Builds on: 3D Gaussian Splatting StegaNeRF Mip-NeRF360
Closest papers by meaning
- SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography (ICLR 2025 · similarity 0.94)
- GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- 3D-GSW: 3D Gaussian Splatting for Robust Watermarking (CVPR 2025 · similarity 0.90)
- GuardSplat: Efficient and Robust Watermarking for 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Geometry Cloak: Preventing TGS-based 3D Reconstruction from Copyrighted Images (NeurIPS 2024 · similarity 0.89)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.88)
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.