Research On JPEG Image Adaptive Steganalysis Based On Deep Learning

With the continuous development of the Internet and information technology,governments,enterprises and individuals rely more and more on digital media such as images and videos for information acquisition and transmission.Steganography can hide some secret information in normal carriers for transmission in order to conceal the existence of secret information.JPEG images are not only the most common image format on the network,but also the most widely used digital media in the process of information transmission.Therefore,JEPG image steganalysis has important strategic significance in maintaining national security,protecting business secrets and personal privacy.This dissertation focuses on JEPG image steganalysis and provides the more effective steganalysis method for JEPG images.At present,the traditional steganalysis methods mostly rely on designing the characteristics artificially.With the application of adaptive steganalysis,the features become more and more complex,and the dimensions are higher and higher,which leads to the increasing difficulty of designing features artificially.In recent years,deep learning has achieved great success in many fields of pattern recognition,and has been gradually applied in image steganalysis.However,the existing steganalysis structures are mostly based on CNN and the network structures are single,which fails to break through the limitation of relatively single scale brought by CNN structure itself.Moreover,most of the current deep-learning steganalysis models are based on CNN structures,and there are few complex hybrid network models that can combine the advantages of different neural networks to apply to steganalysis.This dissertation address the aforementioned problems and contribute the following research work:Firstly,because the residual network can extract more high-dimensional features by deepening the network layers,and the parallel subnet structure enriches the diversity of features,this disseratation designs two composite network structures RCR-CNN and RC-CNN,which combine the residual structure with the parallel subnet structure.Compared with the single CNN network model used for steganalysis in the past,the proposed two composite networks can extract more abundant and diverse high-dimensional features.Secondly,we propose a hybrid network model based on LSTM and CNN,which has two different structures(C-LSTM-1 and C-LSTM-5).C-LSTM combines CNN network with LSTM network,in which we chooses bidirectional LSTM structure.The hybrid network model uses CNN to extract features,and then inputs the extracted features into the LSTM layer for optimization,so that the more effective parts of the original features for steganalysis are remembered,while the unfavorable parts for steganalysis are ignored.Finally,the dissertatioin proves the effectiveness of RCR-CNN and RC-CNN through experiments,and verifies the feasibility and effectiveness of the hybrid network model based on LSTM and CNN through comparative experimental analysis.We improves the RCR-CNN network,which further proves that the network with LSTM structure is more effective for steganography detection.