Research On Image Steganalysis Based On Convolutional Neural Networks

Steganography is a typical information hiding technique that uses digital media,such as texts,images,audios and videos,as carriers to conceal confidential information in a way that is not easily noticeable.Although steganography has played an important role in maintaining information security,it also provides a covert channel for malicious organizations to spread illegal information and other activities.Therefore,to improve the accuracy of steganalysis techniques that can counter steganography,researchers have proposed to use convolutional neural networks to detect image steganography algorithms.By continuously learning the network model parameters and optimizing image feature representation,not only can researchers improve the efficiency of steganalysis but also improve the accuracy of steganography detection.Currently,steganalysis algorithms based on adaptive steganographic images are still relatively single and cannot fully consider the characteristics of adaptive steganography.In addition,manually constructing convolutional neural networks is not only time-consuming and labor-intensive but also may not achieve the desired results.To address these problems,this thesis proposes a new image steganalysis model.A hybrid network model based on convolutional neural network(CNN),Zhu Net-ATTBi LSTM was proposed in this thesis.This method integrates one of the most advanced steganalysis networks,Zhu Net,with attention mechanisms and bi-directional long short-term memory networks(Bi-LSTM).The model first uses convolutional neural networks to extract image features,and then inputs the extracted feature map into the Bi LSTM structure to optimize the connections between features.The processed feature map is then imported into the attention mechanism for further processing.This thesis chooses the optimized model of LSTM,bi-directional LSTM,which allows us to use the sequential learning of forward and backward time steps to fully learn the image feature data.Thus,image feature data can be fully utilized.To address the problem of learning all data indiscriminately in traditional models,the attention mechanism is introduced.Which reallocates the weights of each data to improve the learning of key features,such that it can effectively remove features that are irrelevant to steganalysis and retain features that are effective for steganalysis.At the same time,Zhu Net,which has a high accuracy rate in current steganography analysis models,was selected as the base model to establish the steganalysis model in this thesis.Finally,the improved model was experimentally verified using two publicly available image data sets,Bossbase and COCO.The experimental results show that,this hybrid network model can effectively improve the accuracy of image steganography detection by comparing with previous steganalysis models based on convolutional neural networks.