Mark3DGS: Protecting the Intellectual Property of 3D Gaussian Splatting with Robust Watermarking

Rui Xu, Gaolei Li, Wenkai Huang, Jiazheng Wang, Hang Zhang, Jianhua Li

Keywords: 3D Gaussian Splatting, intellectual property protection, robust watermarking, 3DGS simulator.

Abstract: 3D Gaussian Splatting (3DGS) has become a leading technique in computer vision and graphics, offering photorealistic scene representation and real-time rendering. However, due to high computational demands and the sensitivity of training data, 3DGS models face significant intellectual property theft risks, yet current protection mechanisms are insufficient. In this paper, we introduce Mark3DGS, a novel watermarking framework designed to protect 3DGS models. The framework includes perception-aware pruning for efficient Gaussian reduction, uncertainty-frequency-guided HVQ for resilient watermark embedding, tile-based rasterization with early termination and caching for optimized splatting, and adaptive extraction strategies for reliable watermark recovery. Additionally, we present MarkGS-Sim, a platform to evaluate watermark robustness across various 3DGS variants and conditions. Experimental results show that Mark3DGS outperforms state-of-the-art methods in watermark capacity, imperceptibility, and computational efficiency, achieving 206 FPS rendering, minimal storage ($\textless$ 200MB), compatibility with multiple 3DGS variants, and strong robustness to various watermark attacks.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 10472