神经网络域水印信息优化与加密  

Optimization and encryption of watermarked information in neural network domains

作　　者：马天[1] 赵会敏 杨嫣 杨嘉怡 MA Tian;ZHAO Huimin;YANG Yan;YANG Jiayi(College of Computer Science and Engineering,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学计算机科学与技术学院,陕西西安710054

出　　处：《西安科技大学学报》2022年第3期580-588,共9页Journal of Xi’an University of Science and Technology

基　　金：陕西省自然科学基础研究计划项目(2022JM-508);国家自然科学基金项目(62101432)。

摘　　要：神经网络模型作为一种数字资产,其版权保护日益受重视,针对目前神经网络域水印信息单一、不直观的问题,设计一种水印信息内容优化的方法,在训练神经网络时嵌入水印,采用正则化的方法防止训练神经网络时参数过度拟合,并进行了水印信息的加密研究。通过分析图像域和神经网络域中有效水印算法的需求,将简单的二进制串信息优化为有视觉意义的二值图像与灰度图像,对于不同水印形式进行了分析对比,并在嵌入前对水印信息进行了典型的加密预处理分析,包括Arnold变换、按位异或加密以及行列像素置乱加密。结果表明:该方法可以在不影响原始任务性能的情况下有效地嵌入水印,并且提取的信息质量更好,采用行列像素置乱的二值图像作为水印嵌入,对神经网络性能的影响最小。As a kind of digital asset,the copyright protection of neural network model is paid more and more attention.In order to solve the problem of simple and unintuitive watermarking information in neural network domain,a method of content optimization of watermarking information was proposed.The watermark was embedded in training neural network,the regularization method was used to prevent over-fitting of parameters in the training process,and the encryption of watermark information was studied.The simple binary string information was optimized into visual binary image and gray image by exploring the requirements of effective watermarking algorithm in image domain and neural network domain,the different watermark forms were compared,and typical encryption pre-processing analyses were made of the watermark information before embedding,including Arnold,per-bit dissimilarity encryption and row column pixel scrambling.The experimental results indicate that the method in this paper can effectively embed the watermark without affecting the performance of the original task,and the quality of the extracted information is better.The application of the binary image with row column pixel scrambling in the watermark-embedding will have the least impact on the performance of the neural network.

关 键 词：版权保护 水印信息 神经网络域 图像加密 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]