Research On Image Steganalysis Method Based On LSTM And Reinforcement Learning

With the development of digital multimedia and computer technology,steganography can hide secret information in digital multimedia(such as text,images,audio and video,etc.)in a more imperceptible way.This not only threatens the privacy of every citizen,but also affects the prosperity and stability of the entire country.To improve the accuracy of steganalysis,researchers proposed advanced image adaptive steganalysis algorithms based on convolutional neural networks(CNNs).The algorithms continuously learn the relevant parameters in the network model structure,mine the hidden complex relationships in the data,and extract the data features in the image.This not only greatly reduces the requirements for researchers' experience and energy,but also improves the accuracy of steganalysis.However,the network structures of the existing CNNs-based steganalysis algorithms are relatively monotonous and simple,and does not fully consider the characteristics of image adaptive steganography.At the same time,manually constructing a CNN is a process that takes a lot of researchers' time and machine computing power,and the constructed network is prone to errors.In response to the aforementioned problems,this dissertation proposes two different image steganalysis methods.The main contributions are as follows:Firstly,to further improve the performance of steganalysis,the dissertation proposes a hybrid network model based on CNN and LSTM(Long Short-Term Memory),namely SRNet-res-LSTM.This method uses the residual structure to combine SRNet(which is the most powerful steganalysis network)and LSTM.The model uses a CNN to extract image features,and then transfers the extracted features to the LSTM structure and optimizes the relationship between the features,aiming to retain the effective steganalysis features while ignore the unfavorable features.In the experimental part,we introduced some parameter settings related to LSTM in detail.Experimental results prove that the hybrid network model based on convolutional neural network and LSTM improves the detection accuracy of image adaptive steganography compared with a single CNN.Secondly,in view of the time and computing power that manually constructs the CNN,this dissertation proposes a method to automatically construct the image steganalysis network model to detect image adaptive steganography.The classical reinforcement learning method Q-learning algorithm is used to train the learning agent.We model the action sequence of the agent in the action space as a process of constructing a steganalysis network with a designed module.The agent can search out a variety of high-performance networks suitable for steganalysis from the searching space.We explored the automatic construction of image steganalysis network models,and proposed a module-based searching space with action restrictions,which contains three specific types of modules suitable for steganalysis tasks: noise residual extraction module,feature extraction module,and feature classification module.The use of module-based searching space strategy can effectively reduce the searching space,thereby improves the efficiency of automatically constructing steganalysis networks.We also designed related experiments to prove the effectiveness of the method.The experimental results show that the constructed steganalysis network has performance advantages over the network generated by the state-of-the-art methods automaticly constructing CNNs.