基于可逆神经网络的神经辐射场水印

针对面向隐式表达的神经辐射场的3D模型的版权问题，将神经辐射场水印的嵌入与提取视为一对图像变换的逆问题，提出了一种利用可逆神经网络水印保护神经辐射场版权方案。利用二维图像的水印技术以实现对三维场景的保护，通过可逆网络中的正向过程在神经辐射场的训练图像中嵌入水印，利用逆向过程从神经辐射场渲染出的图像提取水印，实现对神经辐射场以及三维场景的版权保护。但神经辐射场在渲染过程中会造成水印信息丢失，为此设计了图像质量增强模块，将渲染图像通过神经网络进行恢复然后再进行水印提取。同时在每个训练图像中均嵌入水印来训练神经辐射场，实现多个视角下均可提取水印信息。实验结果表明了提出的水印方案，达到版权保护的目的，证明方案的可行性。

Watermarking for neural radiation fields by invertible neural network

Aimin at the copyright problem surrounding 3D models of neural radiation fields focused on implicit representation, this paper tackled this issue by considering the embedding and extraction of neural radiation field watermarks as inverse problems of image transformations. It proposed a scheme for protecting the copyright of neural radiation fields using invertible neural network watermarking. This scheme utilized 2D image watermarking technology to safeguard 3D scenes. In the forward process of the invertible network, the watermark was embedded in the training image of the neural radiation field. In the reverse process, the watermark was extracted from the image rendered by the neural radiation field. This ensured copyright protection for both the neural radiation field and the 3D scene. However, the rendering process of the neural radiation field may result in the loss of watermark information. To address this, the paper introduced an image quality enhancement module. This module utilized a neural network to recover the rendered image and subsequently extract the watermark. Simultaneously, the watermark was embedded in each training image to train the neural radiation field. This enabled the extraction of watermark information from multiple viewpoints. Experimental results demonstrate that the watermarking scheme outlined in this paper effectively achieves copyright protection and highlights the feasibility of the proposed approach.