RDSplat: A Robust Watermarking Framework for 3D Gaussian Splatting Against 2D and 3D Diffusion Editing
RDSplat: Robust Watermarking for 3D Gaussian Splatting Against 2D and 3D Diffusion Editing
AI summary
The authors propose RDSplat, which they describe as the first 3DGS watermarking framework that can simultaneously resist 2D/3D diffusion editing. It uses FAPS to embed a 100-bit watermark into low-frequency Gaussian primitives, replaces the diffusion forward process with Gaussian blur augmentation, and uses a GeoMark decoder to resist editing. Across four benchmarks, seven 2D diffusion attacks and 3D editing, the classic/diffusion robustness at 100-bit capacity is 0.811/0.701, and fine-tuning on an RTX 4090 takes 3 to 7 minutes.