ConcealGS: Concealing Invisible Copyright Information in 3D Gaussian Splatting

TL;DR

ConcealGS addresses copyright protection for 3D-Gaussian Splatting by embedding implicit information into the explicit 3D-GS representation without notably harming rendering quality. It employs a two-stage teacher–student framework with knowledge distillation and a universal decoder, alongside gradient-guided optimization to balance information recovery and visual fidelity. Experiments on NeRF-Synthetic and T&T show ConcealGS delivers robust hidden-information recovery with rendering quality competitive to the original 3D-GS and outperforms NeRF-based steganography across tasks. This work enables practical ownership identification and secure information transfer in 3D content, and suggests directions for extending robustness and applicability.

Abstract

With the rapid development of 3D reconstruction technology, the widespread distribution of 3D data has become a future trend. While traditional visual data (such as images and videos) and NeRF-based formats already have mature techniques for copyright protection, steganographic techniques for the emerging 3D Gaussian Splatting (3D-GS) format have yet to be fully explored. To address this, we propose ConcealGS, an innovative method for embedding implicit information into 3D-GS. By introducing the knowledge distillation and gradient optimization strategy based on 3D-GS, ConcealGS overcomes the limitations of NeRF-based models and enhances the robustness of implicit information and the quality of 3D reconstruction. We evaluate ConcealGS in various potential application scenarios, and experimental results have demonstrated that ConcealGS not only successfully recovers implicit information but also has almost no impact on rendering quality, providing a new approach for embedding invisible and recoverable information into 3D models in the future.

Figures (5)

• Figure 1: llustration of the proposed copyright protection process using 3D-GS. Left: During model training, the owner embeds implicit information into the 3D-GS parameters and uploads the model. Right: During model deployment, when the shared model is used (potentially for commercial purposes), the owner can recover the embedded information from rendered images to identify copyright infringement.
• Figure 2: Overview of ConcealGS: The process involves two key stages. (1) Pre-training: A 3D-GS model is pre-trained as a teacher model for high-quality image rendering. (2) Training: Guided by the teacher model, we simultaneously train a student model and a decoder. The student model learns to render images with embedded implicit information, while maintaining visual similarity to the teacher's output. A gradient-guided optimization strategy, detailed in Section \ref{['sec:2.3']}, is employed to balance rendering quality and information embedding effectiveness.
• Figure 3: Qualitative comparison on NeRF-Synthetic mildenhall2021nerf. Within each column, we show the rendering images on check view and normal view, the recovered hidden image and the residual error compared to the ground truth image. We present the PSNR for scene renderings and recovered hidden images from the check view, along with the PSNR, SSIM, and LPIPS metrics for scene renderings from the normal view.
• Figure 4: Results on the T&T kerbl20233d. Each row displays scene rendering, residual error compared to ground truth image, and the recovered hidden image. We present the PSNR of scene rendering and hidden recovery.
• Figure 5: Robustness Analysis over various (a) Gaussian blur and (b) JPEG compression ratio. SSIM of hidden recovery on T&T kerbl20233d is shown.