RDSplat: Robust Watermarking for 3D Gaussian Splatting Against 2D and 3D Diffusion Editing

3D Gaussian Splatting (3DGS) has become a leading representation for high-fidelity 3D assets, yet protecting these assets via digital watermarking remains an open challenge. Existing 3DGS watermarking methods are robust only to classical distortions and fail under diffusion editing, which operates at both the 2D image level and the 3D scene level, covertly erasing embedded watermarks while preserving visual plausibility. We present RDSplat, the first 3DGS watermarking framework designed to withstand both 2D and 3D diffusion editing. Our key observation is that diffusion models act as low-pass filters that preserve low-frequency structures while regenerating high-frequency details. RDSplat exploits this by embedding 100-bit watermarks exclusively into low-frequency Gaussian primitives identified through Frequency-Aware Primitive Selection (FAPS), which combines the Mip score and directional balance score, while freezing all other primitives. Training efficiency is achieved through a surrogate strategy that replaces costly diffusion forward passes with Gaussian blur augmentation. A dedicated decoder, GeoMark, built on ViT-S/16 with spatially periodic secret embedding, jointly resists diffusion editing and the geometric transformations inherent to novel-view rendering. Extensive experiments on four benchmarks under seven 2D diffusion attacks and iterative 3D editing demonstrate strong classical robustness (bit accuracy 0.811) and competitive diffusion robustness (bit accuracy 0.701) at 100-bit capacity, while completing fine-tuning in 3 to 7 minutes on a single RTX 4090 GPU.