Research And Implementation Of Image Steganalysis Based On Deep Learning

With the development of communication technology,multimedia images have become one of the most important cover for people to transmit information.However,with the emphasis on security issues,multimedia information security has developed rapidly.Information hiding embeds information in the carrier,and then transmits the information through open channels,but this method is likely to be used by criminals to pass some harmful information.Steganalysis can judge the image and determine whether secret information is hidden.Traditional steganalysis is usually a method of manual design with machine learning,which requires a large amount of expert knowledge in feature design,and the results of the machine learning stage cannot guide the design of features,therefore.This paper takes digital images as the research object,and researches the image steganalysis method based on convolutional neural network.The main research results of the paper include:(1)We proposed a steganalysis model based on convolutional neural network Zhu-Net.Compared with the existing CNN-based networks,the advantages of the proposed network focus on:1)The convolution kernel in the preprocessing layer is improved to extract image residuals.At the same time,the convolution kernel is optimized,which reduces the number of parameters and models the local features of the image.2)Use the separable convolution to extract the channel correlation and spatial correlation of the residual features,thereby improving the signal-to-noise ratio and using the image residual features obtained by the preprocessing layer more effectively.3)The spatial pyramid pooling SPP module is used instead of the global average pooling module,and multi-scale data processing can be realized at the same time.Experimental results show that the detection accuracy of the proposed CNN network is significantly better than other networks(2)We proposed a training method based on transfer learning.By using the relevant theories of transfer learning,the steganalysis model can be trained well at low embedding rates,and the models with high embedding rates can be migrated to fine-tuning of low-embedded models speeds up training on low-embedded models and improves performance.The effect of different size datasets on the steganalysis network based on deep learning is studied.By establishing different size datasets and testing on multiple steganalysis networks,experiments show that with data augmentation,the performance of steganalysis network has significant performance improvements.At the same time,the effect of parameter quantity and network operation speed on performance is studied.By using deep separable convolution,the performance of the network is reduced and the inference time is reduced.