Se3DGSMark: Securing Frequency-Based Watermarking with Token Chunking for 3DGS

View at publisher
Authors: Kun Hu, Yibo Zhang, Xiaochao Wang, Hong Zou, Xingjun Wang
Cited by: 1
References: 20

Abstract

Existing 3D Gaussian Splatting (3DGS) watermarking methods often overlook model security and lack robustness against geometric attacks. To address these limitations, we propose a novel end-to-end security framework for 3DGS that integrates key-controlled model access with a frequency-domain watermarking scheme, named Se3DGSMark. First, we enhance overall model security by adversarially optimizing and encrypting the spherical harmonic (SH) coefficients of the 3DGS model, rendering it unusable without a valid key. Second, to ensure robust data embedding, we apply a frequency-domain algorithm to embed watermarking information directly into the rendered image. Furthermore, we embed the watermark exclusively in blue channel, achieving superior imperceptibility with a PSNR exceeding 45 dB. Extensive experiments demonstrate that our method, not only provides a secure mechanism but also surpasses existing methods in both imperceptibility and resistance to geometric attacks.

Fields of Study & Subjects

Advanced Steganography and Watermarking Techniques Biometric Identification and Security Chaos-based Image/Signal Encryption

Publication details

Journal / source
ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Publisher
IEEE
Pages
13497–13501
Publication date
21 April 2026