Rethinking Mesh Watermark: Towards Highly Robust and Adaptable Deep 3D Mesh Watermarking

Xingyu Zhu,Guanhui Ye,Xiapu Luo,Xuetao Wei
2023-12-15
Abstract:The goal of 3D mesh watermarking is to embed the message in 3D meshes that can withstand various attacks imperceptibly and reconstruct the message accurately from watermarked meshes. The watermarking algorithm is supposed to withstand multiple attacks, and the complexity should not grow significantly with the mesh size. Unfortunately, previous methods are less robust against attacks and lack of adaptability. In this paper, we propose a robust and adaptable deep 3D mesh watermarking Deep3DMark that leverages attention-based convolutions in watermarking tasks to embed binary messages in vertex distributions without texture assistance. Furthermore, our Deep3DMark exploits the property that simplified meshes inherit similar relations from the original ones, where the relation is the offset vector directed from one vertex to its neighbor. By doing so, our method can be trained on simplified meshes but remains effective on large size meshes (size adaptable) and unseen categories of meshes (geometry adaptable). Extensive experiments demonstrate our method remains efficient and effective even if the mesh size is 190x increased. Under mesh attacks, Deep3DMark achieves 10%~50% higher accuracy than traditional methods, and 2x higher SNR and 8% higher accuracy than previous DNN-based methods.
Cryptography and Security
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is the lack of robustness and adaptability of existing 3D mesh watermarking methods in the face of various attacks. Specifically: 1. **Robustness problem**: Existing 3D mesh watermarking methods perform poorly in the face of multiple attacks (such as Gaussian noise, rotation, scaling, translation, and cropping, etc.), which easily leads to the loss or damage of watermark information. 2. **Adaptability problem**: Existing methods have inconsistent effects when processing meshes of different sizes and geometric shapes, and cannot effectively deal with unknown mesh categories and larger mesh sizes. To address these problems, the author proposes a new deep 3D mesh watermarking method - DEEP3DMARK. This method improves robustness and adaptability through the following points: - **Using Graph Attention Network (GAT)**: Utilize GAT to embed binary messages in the vertex distribution, making the watermark more robust. - **Adversarial training**: By introducing perturbations of various attacks during the training process, improve the model's resistance to unknown attacks. - **Simplified mesh training**: Use the simplified mesh for training, but be able to maintain effectiveness when dealing with large - scale meshes and meshes of unknown geometric shapes, thereby achieving size and geometric adaptability. ### Specific technical details 1. **Encoding and decoding processes**: - **Encoder**: Take the original mesh \( V \) and the binary message \( M \) as inputs, and generate a watermarked mesh \( \hat{V} \) - **Decoder**: Reconstruct the binary message \( \hat{M} \) from the watermarked mesh \( \hat{V} \) 2. **Loss functions**: - **Encoding loss** \( L_{\text{enc}} \): Minimize the difference between the original mesh \( V \) and the watermarked mesh \( \hat{V} \) to achieve imperceptible embedding. - **Decoding loss** \( L_{\text{dec}} \): Minimize the difference between the reconstructed binary message \( \hat{M} \) and the original binary message \( M \) to achieve accurate reconstruction. - **Discriminative loss** \( L_{\text{dis}} \): Encourage the generated watermarked mesh \( \hat{V} \) to be indistinguishable from the original mesh \( V \) 3. **Experimental results**: - **Robustness**: DEEP3DMARK shows a higher accuracy rate in the face of multiple attacks, which is 10% - 50% higher than traditional methods and 1% - 8% higher than DNN - based methods. - **Adaptability**: DEEP3DMARK can maintain effectiveness when processing meshes of different sizes and geometric shapes. Even when the mesh size is increased by 190 times, it can still maintain an accuracy rate of 81.81%. ### Conclusion DEEP3DMARK significantly improves the robustness and adaptability of 3D mesh watermarking by combining the Graph Attention Network and adversarial training. It can maintain a high accuracy rate under multiple attacks and can effectively process meshes of different sizes and geometric shapes. This provides a new solution for the copyright protection of 3D meshes.