Research On Deep Learning Based Steganalysis In Digital Images

With the rapid development of information technology,illegal applications of steganographies are emerging with increasing frequency,which seriously endangering national and social security.As the steganography detection technology,steganalysis has important practical significance for safeguarding national security and cyberspace security.Deep learning is promoting the research on steganography detection.At present,the steganalysis based on deep learning has surpassed feature-based steganalysis.However,there are still problems to be solved in deep learning based steganalysis.In terms of auxiliary information,the source of auxiliary information for deep learning steganalysis is the image element value or the probabilities with which the image elements are changed during steganographic embedding.However,the auxiliary information is obtained by handcrafted approaches.And the auxiliary information generated by handcrafted approaches are used in a coarse manner as those in features-based steganalysis,which does not fully integrate the characteristics of the deep learning steganalysis.Thus,the steganographic related information in the image is not fully captured and the promotion effect is limited by domain knowledge.In terms of the structure of steganalysis neural network,the complex structure and the large amount of parameter in the steganalysis neural network make the steganalysis neural network over-parameterized,which leads to the existence of redundant or invalid convolution kernel.The invalid or redundant convolution kernel in the deep learning steganalysis limits the distinctiveness of the steganalysis features generated from neural network and the accuracy of steganography detection.Besides,due to the complex structure and the large amount of parameters,the deep learning steganalysis neural network has an over-fitting phenomenon.This results in a decrease in the detection accuracy when there are differences between the steganography algorithm of the training data and the data to be detected in the actual detection scene.Thus,practical application capabilities of the deep learning steganaysis is limited.This dissertation focuses on the problems mentioned above and proposes corresponding countermeasures.The main research work includes the following four aspects.1.Learned-selection-channel-aware steganalytic scheme in spatial domain.The existing selection channels incorporated into spatial steganalysis are designed and used in a coarse manner as those in features-based steganalysis.Besides,the selection channels are not learned and updated during the training,which does not fully exert the potential of deep learning.To solve the problem mentioned above,this dissertation proposes to construct the selection channels in a learning process and present a learned-selection-channel-aware steganalytic scheme in spatial domain.The proposed scheme consists of a selection channel network and a steganalysis network.Both networks are based on CNNs.The two networks are trained together and are unified in one framework optimized by end-to-end.The selection channel network is responsible for learning the selection channels.The steganalysis network takes the images and the learned selection channels as inputs,and outputs the prediction results.The proposed learned selection channels improve the detection accuracy of steganalytic schemes against content-adaptive steganography.In addition,when the source of the training data and the test data are different,or the relative embedding rate of the test data is unknown,the proposed method can still improve the detection accuracy.2.Reference image generation algorithm for JPEG steganalysis based on CNN(Convolutional Neural Network,CNN).To fully capture the the steganographic related information in the JPEG image and improve the detection ability of the deep learning steganalysis for JPEG images,a steganalysis reference image generation method based on convolutional neural networks is proposed.The proposed method generates the reference image from an input image using the convolution layer and the deconvolution layer.And the reference image is taken as auxiliary information for deep learning steganalysis to improve the accuracy of steganography detection.The deep learning steganalysis takes the image and the corresponding reference image as the input data.The proposed method can be pre-trained or trained together with the steganalysis neural network to ensure the reference image generation method consider the characteristics of the JPEG steganalysis neural network.For the detection of content adaptive steganography algorithm,the proposed method can improve the detection accuracy of the existing JPEG steganalysis neural network by 1-6percentage points.3.Convolution kernel reactivation for deep learning steganalysis.The redundant or invalid convolution kernels in the steganalysis neural network limits the feature distinctiveness and affects the accuracy of steganography detection.To solve the problem mentioned above,this dissertation proposes reactivate convolution kernels in steganalysis neural network.The proposed method attempts to reactivate the invalid convolution kernel to enhance the effect of the convolution kernel on steganalysis task and reduce the over-parameterization phenomenon of the steganography detection neural network,make the stegana Accuracy of write detection.First,the effectiveness of the convolution kernel for steganalysis tasks in are measured in terms of the variation in the the convolution kernel weights and the feature distinctiveness.Then,the invalid convolution kernel is reactivated based on filter grafting in the unit of the entire convolution layer.Experimental results demonstrate that the propsed convolution kernel reactivation mthod improves the detection accuracy of deep learning steganalysis.4.Deep representation learning for image steganalysis via constrative learning.The over-fitting phenomenon of the deep learning steganalysis leads to a decrease in the detection accuracy in the case of the embedding method is not contained in training data.To deal with steganographic algorithm mismatch and enhance the generalization ability of the steganalysis neural network,a deep representation learning method for image steganalysis via constrative learning is proposed.The proposed deep representation learning method includes two stages: the pre-training stage and the contrastive training stage.In pre-training stage,the steganalysis neural network is trained via supervised learning.In the contrastive training stage,an image transformation method based on the technique of calibration are proposed.With aid of the proposed image transformation method and steganography algorithm owned by the detector,the images to be detected are converted from unlabled state to labeled state.And positive and negative sample pairs are obtained.In addition,to consider both detection accuracy and generalization ability,the loss function in the contrastive training combines the crossentropy loss function and the similarity between positive and negative samples.The experimental results show that the proposed method can improve the generalization ability of the steganalysis neural network.Focusing on the problems of deep learning steganalysis,this dissertation aims to improve the detection ability of deep learning steganalysis from the aspects of auxiliary information,network parameters and steganalysis feature learning.And the methods proposed in this dissertation can enhance the detection accuracy and generalization ability of deep learning steganalysis.Besides,practicability of deep learning steganalysis is also promoted.