Research On Steganalysis Technology Based On Convolutional Neural Network

With the continuous development and progress of steganographic algorithms,the security and concealment of information hiding are getting higher and higher.More and more steganographic algorithms are applied in information hiding.However,steganography algorithm can be used for copyright authentication,military covert communication and other legitimate purposes,and may also be abused in various illegal activities that destabilize society.Steganography is a double-edged sword,which needs to be used reasonably.The general steganalysis technology is to construct a feature-based image classifier by using some feature changes in the image and supervised learning,and then determine whether the image contains hidden information.However,the traditional general steganalysis technology is facing the problem of more and more complex features and more difficult manual design.The general steganalysis technology based on convolutional neural network is to solve this problem,and it is also one of the hotspots and difficulties in the current steganalysis technology research.In this paper,a general steganalysis model based on convolution neural network is studied.The gray image transform domain and color image transform domain can be detected respectively by utilizing the characteristics of pixel correlation and channel correlation in image domain.The main contributions of this paper are as follows:1)Filter is an important part of convolution neural network,but predecessors have not given how to select filter.The evaluation index of convolution neural network filter core is proposed.The performance of steganalysis model of convolutional neural network depends largely on the filter design,which can help to select the appropriate filter.The experimental results show that the proposed filter evaluation index can guide the filter design and improve the performance of the model.2)Previous research on the combination of convolutional neural network and steganalysis has almost focused on gray image spatial domain.The structure of HCNN,the only steganalysis model in gray image transformation domain,is complex,including quantization,truncation and parallel sub-network structure.Thus,A gray image steganalysis model based on convolution neural network in transform domain is proposed,which is called JPEGCNN.Based on HCNN,this paper optimizes the filter design and simplifies the model structure.The experimental results show that JPEGCNN can detect steganography in transform domain of Jsteg,nsf5,MB1,MB2,J-UNIWARD and so on.Compared with HCNN,JPEGCNN simplifies the structure and reduces the parameters to one twentieth of HCNN,while still maintaining the detection accuracy.3)In reality,most of the images used in the scene are color images.However,the steganalysis model based on convolution neural network for color images has not been studied yet.By studying and designing different feature extraction methods,the steganalysis model JPEGCNN for gray image transformation domain is extended to color image transformation domain,which is called COLOR-JPEGCNN in this paper.COLOR-JPEGCNN subdivision can also be divided.They are RGBMERGE-JPEGCNN model based on RGB three-channel fusion,RGBADD-JPEGCNN model based on RGB three-channel superposition and CHANNEL-JPEGCNN model based on RGB three-channel relationship.The experimental results show that COLOR-JPEGCNN can detect Color-Jsteg,Color-nsf5,Color-MB1,Color-MB2 and Color-J-UNIWARD steganography in transform domain.