Steganalysis Based On Deep Learning

With the development and application of information technology,more attention has been attracted to the communication information security.As one kind of techniques to protect communication security,steganography hides secret information into digital objects,which can be transmitted through public channel without being noticed.However,steganography can also be abused by organizations or individuals for illegal purpose.Steganalysis is rival to the steganography,it is aimed at detecting the existence of steganography.Steganalysis is usually regarded as a binary classification operation to distinguish cover from stego.Traditional steganalysis methods mainly rely on hand-crafted features,which face grate challenges.Aimed at digital image steganalysis,this work focuses on feature learning and classification of steganalysis based on deep learning.The main contributions of this work include the following three folds:(1)A steganalysis framework based on deep learning and difference amplification is proposed.The testing objects are multiplied by a certain coefficient to strengthen the difference.And different kinds of kernels are used to supervise the network learning.Experimental results demonstrate that the proposed method can learn features effectively,and its performance improves a little when compared with SRM and existing steganalysis methods on state-of-the-art spatial steganographic algorithms.(2)A deep learning based framework for adaptive steganalysis is proposed.Weights are assigned adaptively to every pixel during the CNN training process,greater weights are assigned complicated texture area,so the network can learn more matching features.Experimental results on representative modern embedding methods demonstrate the effectiveness of the proposed framework based on CNN.(3)A model of steganalysis based on fully dense connection network is proposed.The dense connection network concatenates all preceding layers outputs as the inputs of the subsequent layers,this connection form increases variation of the subsequent layers inputs.In this work,two improvements are made on original densely connected network:1)Connecting each layer to every other layer throughout the network,and name it fully dense connection network;2)Using an increasing growth rate.Experiments show the effectiveness of fully dense connection network.