TL;DR
WATER-GS introduces an imperceptible, plug-and-play watermarking framework for 3D Gaussian Splatting (3DGS). Instead of altering 3DGS file architectures or relying on fragile image-space pre-processing, it treats the 3DGS scene itself as an implicit encoder, fine-tuning Gaussian attributes against a frozen 2D decoder while simulating 3D-specific point-cloud corruptions.
What Problem Does This Paper Address?
3D Gaussian Splatting (3DGS) has rapidly overtaken Neural Radiance Fields (NeRFs) for real-time, high-fidelity view synthesis. However, securing intellectual property in 3DGS is challenging:
- Explicit representation: Unlike NeRFs, which encode scenes within the implicit weights of a multi-layer perceptron (MLP), 3DGS represents scenes explicitly as millions of colored, oriented 3D ellipsoids.
- Failure of intuitive baselines: Pre-watermarking training images fails because the splatting and $\alpha$-blending rendering process scrambles 2D watermark patterns (yielding ~50% bit error rates, equivalent to random guessing). Conversely, appending watermark bits directly into explicit 3D attributes (such as Spherical Harmonics) severely degrades visual fidelity.
- 3D transmission distortions: Real-world 3D assets undergo pruning, point-cloud compression, and coordinate jitter, which degrade watermark signals embedded within raw scene geometry.
Methods
WATER-GS adopts a plug-and-play strategy inspired by fixed neural network steganography:
[Pre-trained 3DGS] ──► [3D Distortion Layers (GN, Dropout, Crop)] ──► [Differentiable Splatting] ──► [Rendered Image] ──► [Frozen Decoder D] ──► Extracted Bits
- Offline Universal Decoder: A 2D convolutional encoder-decoder pair (based on HiDDeN) is pre-trained end-to-end on natural images using Mean Squared Error (MSE) and Binary Cross Entropy (BCE) message loss. Once trained, the decoder $D$ is frozen and reused across any 3DGS model.
- Scene-Level Fine-Tuning: Rather than modifying network architectures or adding explicit 3D watermark encoders, the original 3DGS optimization loop acts as the encoder. The Gaussian parameters $\Theta = \{\mu, s, r, \alpha, h\}$ (position, scale, rotation, opacity, spherical harmonics) are fine-tuned to simultaneously minimize standard RGB photometric losses ($L_1 + \text{SSIM}$) and the BCE loss between a target $l$-bit message and the message decoded from 2D rendered views: $$\mathcal{L}_{tot} = \mathcal{L}_{rgb} + \gamma \mathcal{L}_m$$
- 3D Distortion Layers: During fine-tuning, differentiable 3D distortion layers perturb the Gaussian parameters prior to rendering:
- Gaussian Noise (GN): Perturbs coordinate positions ($\sigma = 0.01$).
- Dropout: Randomly removes a fraction $p = 10\%$ of Gaussians to build resilience against pruning.
- Crop: Drops contiguous blocks of points to simulate partial geometric loss.
Datasets
- Decoder Training: MS-COCO (Lin et al., 2014) is used for training the universal 2D watermark decoder.
- 3DGS Evaluation: Evaluated across standard novel-view synthesis benchmarks:
- LLFF (forward-facing scenes)
- Mip-NeRF 360 (unbounded indoor and outdoor scenes)
- Tanks & Temples (Truck and Train scenes)
Key Insights and Ideas
- 3DGS as an Implicit Encoder: Because 3DGS is over-parameterized (millions of Gaussians), subtle perturbations distributed collectively across positions, opacities, and color harmonics can reliably project a watermark into any 2D camera view without introducing visible image artifacts.
- Joint Attribute Optimization is Key: Ablations show that optimizing only positions ($xyz$) or only spherical harmonics ($h$) leads to either incomplete watermark decoding (BER $> 10\%$) or noticeable color fringing. Distributing message gradients across all parameters achieves the best balance between extraction accuracy and rendering fidelity.
- Modeling 3D Degradation Directly: Prior 3D watermarking methods focus heavily on 2D image-space degradations (e.g., JPEG compression). Introducing 3D point perturbations during training provides robustness against downstream mesh/point pruning and file compression.
Key Results
- Robust Extraction: Using 48-bit messages, WATER-GS achieves a Bit Error Rate (BER) of 3.26% on LLFF, 9.22% on Mip-NeRF360, and 6.30% on Tanks & Temples under uncorrupted (Identity) settings. Baselines like HiDDeN+3DGS and MBRS+3DGS fail completely (BER $\approx 47\text{--}54\%$).
- Distortion Layer Gains: Under 3D Gaussian noise attacks, fine-tuning with 3D distortion layers improves extraction accuracy by over 20% compared to fine-tuning without them (BER drops from $25.62\%$ to $4.86\%$).
- Outperforming NeRF Baselines: Compared to CopyRNeRF (8.84% BER on 16 bits), WATER-GS achieves 2.83% BER with less degradation to image quality, while supporting higher payload capacities (48 bits).
- Speed & Broad Compatibility: Fine-tuning takes 45 to 100 minutes on a single RTX 3090. The method applies directly to 2D Gaussian Splatting (2DGS, BER 8.53%) and compressed representations like Compact3D (BER 9.85% after point compression).
Limitations and Future Work
- Training Overhead: Incorporating 3D distortion operations incurs a noticeable slowdown during fine-tuning due to slow tensor replacement routines.
- Sensitivity to Extreme Noise: Severe geometric noise in complex outdoor environments remains challenging (e.g., BER rises to $29.11\%$ under Gaussian noise on Tanks & Temples).
- Visual Trade-off: High-capacity embedding causes a mild drop in fidelity ($\sim 1.5\text{--}2.2$ dB PSNR reduction). Future directions include refining perceptual weighting to preserve visual quality and enhancing robustness against non-rigid 3D transformations.
Summary updated