WATER-GS: Toward Copyright Protection for 3D Gaussian Splatting via Universal Watermarking

TL;DR

This work tackles copyright protection for 3D Gaussian Splatting (3DGS) by introducing WATER-GS, a universal watermarking framework that embeds imperceptible watermarks into 3DGS while preserving rendering quality. It combines a pre-trained universal watermark decoder with a watermark-embedding pipeline that fine-tunes the 3DGS parameters to carry an $l$-bit message, augmented by 3D distortion layers to simulate real-world 3D data distortions. Empirical results show WATER-GS achieves robust watermark extraction across diverse 3DGS variants and distortions, outperforming NeRF-based baselines, with up to ~20% improvement in extraction accuracy and compatibility with 2DGS and compression pipelines. This approach provides a practical, scalable solution for ownership provenance in 3D content and can adapt to evolving 3DGS formats and workflows.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a pivotal technique for 3D scene representation, providing rapid rendering speeds and high fidelity. As 3DGS gains prominence, safeguarding its intellectual property becomes increasingly crucial since 3DGS could be used to imitate unauthorized scene creations and raise copyright issues. Existing watermarking methods for implicit NeRFs cannot be directly applied to 3DGS due to its explicit representation and real-time rendering process, leaving watermarking for 3DGS largely unexplored. In response, we propose WATER-GS, a novel method designed to protect 3DGS copyrights through a universal watermarking strategy. First, we introduce a pre-trained watermark decoder, treating raw 3DGS generative modules as potential watermark encoders to ensure imperceptibility. Additionally, we implement novel 3D distortion layers to enhance the robustness of the embedded watermark against common real-world distortions of point cloud data. Comprehensive experiments and ablation studies demonstrate that WATER-GS effectively embeds imperceptible and robust watermarks into 3DGS without compromising rendering efficiency and quality. Our experiments indicate that the 3D distortion layers can yield up to a 20% improvement in accuracy rate. Notably, our method is adaptable to different 3DGS variants, including 3DGS compression frameworks and 2D Gaussian splatting.

Figures (8)

• Figure 1: Before publication, creators can utilize pre-trained decoder and 3D distortion layers to embed digital signatures as watermarks into 3DGS files, thereby asserting ownership. Even when the 3DGS files undergo various distortions, the watermark can be reliably extracted, serving as effective proof of ownership.
• Figure 2: Two intuitive solutions for 3DGS watermarking. Watermarking Training Datasets draws inspiration from existing model watermarking techniques, leveraging watermarked datasets to embed watermarks. Concatenating Watermarks with Attributes incorporates an explicit encoder to seamlessly integrate watermarks with specific attributes.
• Figure 3: Illustration of our WATER-GS framework. (1) Training a universal decoder. In this stage, a message encoder ${\mathcal{E}}$ and a decoder ${\mathcal{D}}$ are trained end-to-end. (2) Fine-tuning the 3D Gaussians. In this stage, creators fine-tune the original 3D Gaussians using the pre-trained decoder to embed fixed messages. They have the flexibility to control which parameters are frozen by employing a self-defined mask. (3) Distorstion layers. The 3D distortion layers ${\mathcal{N}}$ introduce alterations to the 3D Gaussians $\Theta$, resulting in a distorted $\hat{\Theta}$.
• Figure 4: Visual quality comparisons of each baseline. We display both the rendered images and the corresponding residual images ($\times$ 10). WATER-GS demonstrates the optimal balance between rendering quality and watermark extraction accuracy.
• Figure 5: A visual comparison of the watermark patterns pre- and post-rendering. The residual images indicate that the watermark pattern is disrupted.