Research On A Universal Image Steganalysis Algorithm Based On Deep Learning

In the information society,image is one of the commonly used steganography carriers,and illegal organizations also use steganography technology to transmit messages,which has laid a hidden danger for national security and social security.In order to maintain national security,many scholars are committed to steganalysis.Image steganalysis is to find the possibility of image feature modification by analyzing the characteristics of the sample image,and then judge whether the sample carries secret information.In order to improve the accuracy of steganalysis,using deep learning to design steganalysis algorithm has become one of the current research hotspots.The traditional steganalysis method is to extract the image features manually,and finally classify the samples by training the feature classifier,so as to determine whether the sample image is a secret image.The steganalysis algorithm based on deep learning integrates the extraction of image features,the training of classifier and the discriminative output of classifier into the deep learning network,so the accuracy and efficiency of classification are higher.Most of the existing steganalysis algorithms based on deep learning design the analysis network for the steganalysis algorithm in a certain domain of the image,but the domain in which the sample to be tested is embedded in the secret message is unknown in the actual operation,so it is often necessary to use multiple steganalysis algorithms for different domains to obtain the higher confidence rate of the discrimination results,which leads to the actual steganalysis workload,The confidence level is not high.Therefore,in order to reduce the workload and improve the efficiency of steganalysis,this paper designs and constructs a universal steganalysis algorithm based on deep learning.The algorithm uses parallel architecture,and improves and optimizes yenet,so that the deep learning network has better steganalysis ability for steganalysis algorithms in various domains,and can achieve higher results through one steganalysis The discriminant confidence rate of the model is given.The main work of this paper is as follows1.In order to improve the generality of steganalysis algorithm based on deep learning,by studying the relationship between spatial pixel value of image and quantized DCT coefficient in JPEG domain,it is noticed that spatial pixel distribution has some correlation with DCT coefficient distribution in JPEG domain,so the network is improved.The optimized main network is mainly used to learn spatial pixel distribution characteristics,and a parallel sub network based on JRM(JPEG rich model)is added to learn the distribution characteristics of quantized DCT coefficients.Finally,the generated feature graph is combined with the main network feature graph by concat function.The experimental results show that the parallel sub network can improve the generality of the algorithm,and the algorithm has good accuracy for steganalysis in both spatial domain and JPEG domain.2.In order to improve the accuracy and efficiency of the algorithm,by studying the relationship between the full connection layer and the global average pooling layer,yenet is optimized,and the full connection layer in yenet is replaced by the global average pooling layer,so the probability of over fitting problem in the case of small and medium data sets is reduced,which is conducive to improving the efficiency and accuracy of the network,and the confidence rate of classification is higher.By selecting some images from the public image data set as the data set,the performance of the steganalysis algorithm constructed in this paper is tested and compared,and the steganalysis results are analyzed.Experimental results show that the proposed network has good performance in both spatial domain and JPEG domain.Therefore,the steganalysis network is universal,and the overall time-consuming of steganalysis is more improved than the original algorithm,which makes the steganalysis algorithm in this paper have higher efficiency.