Research On Digital Image Steganography Algorithm Based On Deep Learning

Steganography is a kind of technology that hides information into a carrier without suspect.Information is disguised as a public carrier,which greatly reduces the possibility of malicious interception of information,so that information is transmitted more securely.It plays an important role in the field of information transmission and information security.With the development and popularization of digital media technology,the use of steganography is becoming more and more frequent.Therefore,steganography has become one of the research hotspots.Digital image steganography is a very popular research direction for steganography because of abundant carrier images.In this thesis,a new digital image steganography algorithm is proposed.Firstly,the steganographic ability of carrier image is estimated,then the carrier image is regenerated according to steganographic ability,and finally information is embedded using a specific steganography algorithm.The main work and contributions of this thesis are as follows:(1)The relationship between the parameter,of carrier image and the steganographic ability of carrier image is found,after studying the process of embedding messages.A A estimation algorithm based on neural network is designed.ImageNet dataset are validated using the 0.4bpp-payload HILL(HIgh-pass,Low-pass,Low-pass)algorithm(with SRM+EC detector,Spacial Rich Model+Ensemble Classifer)and the experimental results are as follows:the images with λ value in the[0,10]interval,which are unsuitable to embed,are steganized,and the error rate is 0.084,the images with λ value in the[50,60]interval,which are suitable to embed,are steganized,the error rate is 0.4401.In ImageNet test set,the correlation coefficient between λ and the output obtained by the λ estimation algorithm based on neural network is as high as 0.99984.(2)A neural network based carrier image generation algorithm is proposed,which can generate images of any size.There are two branches in the generation network,one branch is used for detail generation,the other branch does nothing,and the outputs of two branches are merged into one output image at the end of the network.Modification probability calculation network,message embedding network and steganographic detection network are used to train generating network,and the output of steganographic detection network guides the convergence of carrier image generation network.Using the 0.4bpp HILL algoritlm as stego algorithm(with SRM+EC detector),the error rate of the detection algorithm increases by 4%after regenerating whole BOSSBASE database.(3)A spatial domain digital image steganography algorithm based on carrier evaluation and carrier generation is proposed.Whether a carrier image needs to be regenerated is due to the λ of carrier images.Carrier image withλ less than the threshold is regenerated and the message is embedded with a specific steganography algorithm.The algorithm does not depend on a specific embedding algorithm,and the embedding algorithm can be selected accord-ing to the actual situation.Using 0.4bpp Hill algorithm for validing under the BOSSBase dataset,regenerating 8000 images can increase the error rate by 4 percent.