SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography

TL;DR

SecureGS tackles privacy and integrity challenges in 3D Gaussian Splatting by introducing an anchor-point–driven architecture that decouples original and hidden content through Hybrid Decoupled Gaussian Encryption Representation. A region-aware density optimization strategy confines hidden information to regions that minimize geometric leakage while maintaining rendering speed and fidelity. The approach yields higher quality renderings, reduced storage, and stronger security than prior 3DGS steganography methods, and extends to hiding 3D objects, 2D images, and bits. This has practical impact for copyright protection and secure transmission of complex 3D assets in real time.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a premier method for 3D representation due to its real-time rendering and high-quality outputs, underscoring the critical need to protect the privacy of 3D assets. Traditional NeRF steganography methods fail to address the explicit nature of 3DGS since its point cloud files are publicly accessible. Existing GS steganography solutions mitigate some issues but still struggle with reduced rendering fidelity, increased computational demands, and security flaws, especially in the security of the geometric structure of the visualized point cloud. To address these demands, we propose a SecureGS, a secure and efficient 3DGS steganography framework inspired by Scaffold-GS's anchor point design and neural decoding. SecureGS uses a hybrid decoupled Gaussian encryption mechanism to embed offsets, scales, rotations, and RGB attributes of the hidden 3D Gaussian points in anchor point features, retrievable only by authorized users through privacy-preserving neural networks. To further enhance security, we propose a density region-aware anchor growing and pruning strategy that adaptively locates optimal hiding regions without exposing hidden information. Extensive experiments show that SecureGS significantly surpasses existing GS steganography methods in rendering fidelity, speed, and security.

Figures (14)

• Figure 1: Analysis of previous 3DGS steganography method GS-Hider zhang2024gs.
• Figure 2: Overall Pipeline of our SecureGS. 3D objects, images, and bits can be hidden in the original 3D scene, and only authorized users can decode these hidden messages. The core of our method is to ensure both the file format and geometric structure security of the public anchor points.
• Figure 3: Overall framework of our SecureGS. We first voxelize the scene, where each voxel contains an anchor point with the position $\mathbf{x}_v$, feature $\boldsymbol{f}_v$, scaling factor $\boldsymbol{l}_v$, and offsets $\{\boldsymbol{O}_{v\circledast i}^{ori}\}_{i=1}^k$. Then, we explicitly compute the positions $\boldsymbol{\mu}_{v\circledast i}^{ori}$ via Eq. \ref{['ori_position']}, and predict attributes $\{\boldsymbol{c}_{v\circledast i}^{ori}, \alpha_{v\circledast i}^{ori}, \boldsymbol{q}_{v\circledast i}^{ori}, \boldsymbol{s}_{v\circledast i}^{ori}\}$ via a series of public MLPs. Meanwhile, we use private offset predictor $\mathcal{F}_o^{\dagger}$ and MLPs to store the position and attributes of the Gaussian points representing the hidden object. Finally, we design a region-aware density optimization to control the Gaussian point growing and pruning.
• Figure 4: Visualization of the point cloud produced by our SecureGS without region-aware density optimization (RDO) strategy. The RGB reference of the hidden scene is placed on the left bottom.
• Figure 5: Rendering quality comparison of the hidden message between the proposed SecureGS, previous GS-Hider, and 3DGS+StegaNeRF. For our SecureGS, we also present the decoupled original scene and hidden object on the $4^{th}$ and $5^{th}$ row, which cannot be achieved by other methods.