Research On Steganalysis Based On Convolutional Neural Networks

Steganography conceals secret information in public covers,so as to achieve the purpose of covert communication.The illegal use of steganography has posed a serious threat to national security and military safety.The research of steganalysis is of great importance to curb the illegal use of steganography.Based on convolutional neural network,a typical deep learning model,steganalysis technology is studied.The main work of this article is as follows:1.On the basis of analyzing and summarizing the basic ideas of deep learning,the structure of convolutional neural network and the training method of convolutional neural network,the framework of steganalysis under convolution neural network is proposed.The key points in the framework include: dividing image set,making structure of nueral network,training neural network and steganalysis based on neural networks.Dividing image set means dividing existing image set into training set and validation set.Training set is used to training networks whereas validation set is used to verify the performance of current network during training.Making structure of nueral network meanings the structure of the convolutional neural network is constructed according to the characteristics of the steganalysis task.The structure includes the preprocessing layer,the processing layer for the steganalysis,the conventional convolution module,the full connection layer and the Softmax layer,etc.Training neural network means according to the objective function and specific training algorithm,the weight parameters in the network can be optimized to make the network decision more precise.Steganalysis based on neural networks indicates steganalysis based on convolution network should be divided into two cases: same source and different source.In the same source case,the strategy of random subspace processing and sub region joint decision is applied to steganalysis.In the different source case,steganalysis is implemented by generating matching images,convolution network feature extraction and other strategies.2.In order to deal with the problem that the existing steganalysis methods based on deep neural network is not closely integrated with steganographic characteristics,a steganalysis method based on convolutional network with random subspace is proposed.The scheme focuses on the design of convolution network structure based on random subspace,and steganalysis is directly implemented by the designed network.In the network structure,the image is preprocessed using the extraction method of rich model features.The processing first uses a variety of high pass filter templates to process images to obtain a variety of residual graphs,then calculates the symbiotic matrix of the residual graph,and combines multiple symbiotic matrix data into the rich model data.This process forms pre-processed layer of the deep convolution neural network.It is concluded that the processing of the FLD ensemble classifier is equivalent to a three layer neural network,and the FLD ensemble classifier is modeled by the convolution neural network.On this basis,a random subspace processing module is designed.The module,as the shallow part of the network,includes a Randomdata layer,a convolutional layer and a ReLU nonlinear activation layer;further deepen the network structure,expand the deep processing module;using a classification module as the output module of the network,which includes a full connection layer and a Softmax layer.3.In order to deal with the problem that the existing steganalysis methods based on deep neural network can not detect adaptive steganographies well,a convolutional network steganalysis method based on sub region joint decision is proposed.The analysis shows that the difference between the steganalysis task and the computer vision task is that any image subregion can form an effective sample.Based on this analysis,a single image is divided in order to construct a variety of subsample images.The sub sample images form multiple sub decisions by the after-trained deep convolutional neural network,and the final decision results are obtained by the joint decision of majority voting method for all sub decision making.The diversity of subsample images ensures the enhancement of the accuracy of the joint decision.In order to improve the time efficiency of the proposed method,an analysis is made that any image subregion can be mapped into a subregion in the map of each layer of the network.Based on this analysis,a parallel acceleration strategy based on the full convolution network for decision is proposed.The full convolution network changes the last full-connected layer in the network used in training phase to convolutional layer,and the remaining layers remain unchanged.4.In order to deal with the problem that the existing steganalysis methods based on deep neural network have not considered problem of cover source mismatch(CSM),steganalysis with convolutional nueral network faced with cover source mismatch is studied,and a anti-CSM method based on convolutional neural network is proposed.The analysis shows that the reason that the accuracy of the existing method is reduced is the interference of the image content,and the existing residual filtering method can not completely eliminate the interference of the image content to the steganalysis.The proposed method is based on the convolution neural network after training.In the decision,the Gauss low pass filter is used to process the testing image,to remove the possible stego signal and to produce a match image with similar image content to the image to be judged.The steganography is applied to the matching image using the recognized steganography algorithm,and the stego of the match image is obtained.After the testing image,the match image and the stego of match image going through the trained network,the values of the second layers of the network are taken as the feature of steganalysis.The similarity of the matching feature and the testing feature and the similarity of the stego match feature and the testing feature are calculated respectively.The final decision is obtained according to the two similarity.Two similarity computation methods are proposed,which are inner product method and weighted variance method.The proposed steganography can not only be used in steganalysis,but also can be used in steganography recognition.In the task of steganography recognition,algorithms unknown in training can also be recognized.