Research On Deep Learning Based Watermarking Attack And Defense Technology

With the rapid development of digital technology and the increasing popularity of the Internet,digital images,audio and video,as important carriers of information,have gradually become indispensable factors in daily life and can be seen everywhere in social platforms,commercial advertisements and film and television industries.Among them,a considerable part of digital information has high protection value due to cost,use and other reasons.As an effective means of property rights protection and disclosure tracing,digital watermarking technology has been gradually paid attention to.More and more film and advertising manufacturers choose to adopt appropriate digital watermarking technology to protect their products.At the same time,some manufacturers want to check whether some digital information contains copyright information before commercial use.If the test results show that the content is protected,manufacturers can avoid copyright disputes by applying for authorization or giving up the use of such information.This topic comes from the watermarking attack and defense cooperation project.The actual demand has two main starting points,one is to embed the watermark to protect our own data,the other is to detect the watermark of the third party data,so as to avoid the infringement and disputes caused by the use of the third party data.In terms of copyright protection,this paper adopt the method of calibration template with embedded algorithm combination,designed a comprehensive watermark embedding scheme,including calibration template based on Fourier spectrum image design,using its some excellent properties,solve the problem of watermark image decoding the heavy synchronization,can accurately restore watermark image of geometric attacks;The embedded algorithm adopts deep learning to ensure the robustness of the algorithm to JPEG compression from two aspects of network structure and training mode.In terms of watermark detection,this paper designs a stepwise steganographic analysis convolutional neural network based on deep learning technology,and further improves the network detection accuracy with the help of existing optimization methods.The algorithm achieves over 99% detection accuracy on the target data set,and shows good generalization performance on some data sets.