Robust 3D Gaussian splatting watermarking via hybrid optimization

Abstract

3D Gaussian Splatting (3DGS) has demonstrated substantial commercial potential due to its excellent real-time rendering, thus increasing the demand for digital watermarking for copyright protection. However, the rendering of 3DGS critically depends on millions of Gaussian points within the model. These points are vulnerable to noise interference and malicious tampering, making it difficult for existing methods to achieve an optimal balance between watermark robustness, rendering fidelity, and storage efficiency. To overcome these limitations, we propose HybridGSW, a robust watermarking framework using hybrid group-instance optimization. First, an importance evaluation and pruning of Gaussian points are conducted within the pre-trained 3DGS model. Subsequently, a non-local Gaussian grouping strategy is introduced based on attribute similarity. Based on this, a hybrid optimization mechanism is implemented by simultaneously performing group-level consistent optimization and instance-level fine-grained optimization. This approach effectively combines the robustness of group-level non-local redundant watermark embedding with the fidelity preservation of instance-level refinement. Experimental results demonstrate that HybridGSW outperforms state-of-the-art (SOTA) methods in both rendering quality and watermark robustness.