Levenberg-Marquardt deep neural watermarking for 3D mesh using nearest centroid salient point learning

Area of Science:

• Computer Vision
• Information Security
• Digital Watermarking

Background:

• 3D mesh models require robust watermarking for protection against unauthorized use.
• Existing methods face challenges in maintaining imperceptibility and resisting diverse attacks.

Purpose of the Study:

• To propose a novel watermarking authentication method (NCDG-LV) for 3D mesh models.
• To enhance distortion detection and recovery capabilities.
• To improve robustness against common signal processing and geometric attacks.

Main Methods:

• Salient point detection using the Nearest Centroid and Discrete Gaussian Geometric (NC-DGG) model.
• Map segmentation of 3D mesh models based on salient points.
• Watermark embedding using the Multi-function Barycenter.
• Watermark extraction and authentication via Levenberg-Marquardt Deep Neural Network and back-propagation.

Main Results:

• The NCDG-LV method demonstrates high imperceptibility and tolerance to attacks like smoothing, cropping, translation, and rotation.
• Achieved superior performance in salient point detection time, distortion rate, true positive rate, PSNR, BER, and RMSE compared to state-of-the-art methods.

Conclusions:

• The proposed NCDG-LV method offers an effective solution for secure 3D mesh watermarking.
• It provides robust distortion detection, recovery, and authentication with high fidelity.