Abstract
As 3D Gaussian Splatting (3D-GS) gains significant attention and its commercial usage increases, the need for watermarking technologies to prevent unauthorized use of the 3D-GS models and rendered images has become increasingly important. In this paper, we introduce a robust watermarking method for 3D-GS that secures copyright of both the model and its rendered images. Our proposed method remains robust against distortions in rendered images and model attacks while maintaining high rendering quality. To achieve these objectives, we present Frequency-Guided Densification (FGD), which removes 3D Gaussians based on their contribution to rendering quality, enhancing real-time rendering and the robustness of the message. FGD utilizes Discrete Fourier Transform to split 3D Gaussians in high-frequency areas, improving rendering quality. Furthermore, we employ a gradient mask for 3D Gaussians and design a wavelet-subband loss to enhance rendering quality. Our experiments show that our method embeds the message in the rendered images invisibly and robustly against various attacks, including model distortion. Our method achieves superior performance in both rendering quality and watermark robustness while improving real-time rendering efficiency.
Introduction
3D Gaussian Splatting (3D-GS) has rapidly become the industry standard for real-time, photo-realistic 3D reconstruction. However, its rise has brought a significant challenge: the risk of "model piracy" and unauthorized image use. While digital watermarking is well-established for 2D images and has been adapted for Neural Radiance Fields (NeRFs), 3D-GS presents unique structural challenges.
A typical 3D-GS scene consists of millions of individual Gaussians. Many of these are redundant or contribute minimally to the final render. Embedding watermarks into these low-impact Gaussians often results in fragile signatures that are easily lost during model editing or image compression. Furthermore, existing NeRF-based watermarking methods are computationally prohibitive, often requiring hours for embedding. 3D-GSW addresses these issues by providing a robust, invisible, and efficient watermarking framework.
Methodology
The 3D-GSW framework introduces three core innovations to ensure watermark robustness and model efficiency:
1. Frequency-Guided Densification (FGD)
FGD is a two-phase preprocessing step designed to optimize the model before embedding.
- Phase 1: Pruning. The system identifies and removes "weak" Gaussians that have a negligible impact on rendering quality. Counter-intuitively, removing nearly 30% of the Gaussians makes the watermark more robust by forcing the message into the most "important" structural components of the scene.
- Phase 2: High-Frequency Splitting. Using a Discrete Fourier Transform (DFT), the system identifies high-frequency areas (textures and edges) and splits Gaussians in these regions. This creates a more efficient structure that can "hold" a watermark in areas where the human eye is less sensitive to subtle changes.
2. Gradient Masking
To maintain visual fidelity, an exponential gradient mask is applied during fine-tuning. This mask restricts the degree to which Gaussian parameters—such as position, color, and scale—can be modified, ensuring the watermark is embedded with high precision without introducing visual artifacts.
3. Wavelet-Subband Loss
During training, a Discrete Wavelet Transform (DWT) analyzes the rendered images. By focusing on high-frequency subbands, the loss function ensures the watermark does not blur fine details or introduce "ghosting" effects, preserving the photo-realism of the 3D-GS model.
Experimental Results
Performance and Accuracy
3D-GSW was evaluated across standard benchmarks including Blender, LLFF, and Mip-NeRF 360. The method achieved a bit accuracy of 97.37% for 32-bit messages, significantly outperforming baselines such as StegaNeRF (93.15%) and WateRF (93.42%).
Robustness and Efficiency
The watermark proved resilient against aggressive distortions, including 50% JPEG compression and 40% image cropping. It also survived "model attacks" where 20% of the Gaussians were randomly removed or cloned.
Notably, 3D-GSW presents an "Efficiency Paradox": while most watermarking schemes add overhead, 3D-GSW improves the model. Due to FGD pruning, watermarked models are approximately 23% smaller in storage and render 28% faster (increasing from ~56 FPS to ~72 FPS) compared to original un-watermarked models. Furthermore, the embedding process is exceptionally fast, completing in just 21 minutes, compared to the 7 hours required by previous state-of-the-art methods.
Conclusion
3D-GSW provides a robust solution for copyright protection in the era of real-time radiance fields. By leveraging frequency-domain analysis and intelligent pruning, it achieves high bit accuracy and robustness without sacrificing—and indeed, while improving—rendering performance. Future work will explore expanding the system to embed more complex data types, such as encrypted metadata, directly into the 3D radiance field.
Summary updated