Study On Robust Video Zero_watermarking Based On Convolutional Neural Network

With the rapid development of mobile internet technology and the rise of various short video-sharing applications,the number of online videos has increased dramatically,and the spread of the video has also become wider.Some security issues have attracted lots of attention.There are many pirated videos on the Internet,research on the methods of digital video copyright protection is particularly important in this situation.Digital watermarking technology is an effective solution for video copyright protection and content authentication.However,in the case of high visual quality requirements,the existing robust video watermarking methods are difficult to ensure both imperceptibility and robustness,and it is also unable to effectively resist the desynchronization attacks.This paper combines convolutional neural network(CNN),Polar Complex Exponent Transform(PCET)in the field of signal processing,chaotic image encryption algorithm and various techologies to propose a robust video zero-watermarking scheme which can effectively resist desynchronization attacks.The specific work of this project is as follows:(1)In this thesis,a robust video zero-watermarking scheme for copyright protection using the combination of visual geometry group network(VGGNet),self-organizing map(SOM)and PCET is presented.The scheme can not only make up for some existing problems like lacking performance in evaluating,but also enhance the robustness.Firstly,VGGNet is utilized to extract the content features of each frame,which is used as the input of SOM for cluster analysis,and the key frame of video is selected by the maximum entropy value.Secondly,the PCET is applied to abstract invariant moments of key frames,and further,is scrambled through chaotic mapping,and dimension is reduced by singular value decomposition(SVD).Next,by comparing the adjacent values of the obtained maximum singular value sequence,a binary sequence is generated and scrambled.Finally,an exclusive-OR(XOR)operation is imposed on the scrambled binary sequence and the chaotic logic map encrypted watermark to generate a zero-watermark signal,and registered in a third-party public database for the inspection use.Simulation experiment results show that the zero-watermark signal generated by the proposed method has sufficient equilibrium and distinguishability.In resisting common signal processing attacks,especially for desynchronization attacks such as geometry and inter-frame,the NC value is greater than 0.9 which is robust.In addition,compared with the existing video zero-watermarking and traditional robust video watermarking methods,the proposed scheme exhibits superior robustness.(2)This thesis proposes a robust video zero-watermarking algorithm based on residual neural network(ResNet),2D sine logistic modulation map(2D-SLMM)and PCET domain.First,the original video is shot-divided by a correlative coefficient-based camera lens segmentation algorithm,and the frame with the largest information entropy in each shot is selected as the key frame.Then,the invariant moments of key frames are extracted by PCET as the input of ResNet network,and the output result of ResNet is extracted as the video content feature sequence and scrambled.Next,the binary feature sequence is obtained by comparing the absolute value of each single value with the overall mean in the scrambled feature sequence.Finally,the zero-watermark signal is generated by the XOR operation on the obtained scrambled binary sequence and the mixed chaotic encrypted watermark,and registered in a third-party public library for checking use.Simulation experimental results show that the zero-watermark signal generated by this method can not only maintain good equilibrium and distinguishability,but also improve the robustness greatly,the NC values are all greater than 0.95.Especially,when resisting the interframe attacks,the NC values can achieve 0.99.And it shows certain advantages when compared with other watermarking methods.This dissertation proposes two robust video zero watermarking algorithms,whose purpose is to protect video copyright more effectively when the visual quality requirements are high or faced with geometric attacks within frames,inter-frame(frame loss,frame averaging and frame swapping)and other desynchronization attacks.