Research On Image Steganalysis And Hidden Information Reconstruction Based On Deep Convolutional Neural Network

Image steganalysis is the process of detecting the presence of hidden information in an image.Existing image steganalysis methods cannot fully extract hidden image information even if the presence of hidden information is successfully detected,resulting in either poor detection performance or inability to reconstruct hidden information.The deep learning steganalysis algorithm can obtain better detection performance and better generalization ability through training based on a large amount of data,which is very suitable for practical applications in real scenes.In this paper,the proposed image steganalysis algorithms based on deep learning not only classifies stego and carrier images,but also reconstructed secret image from stego images.The main research achievements are as follow:(1)SG-Res Net: Steganalysis Rich Model and Gabor Residual Network for Color Image Steganalysis.At present,there are few steganalysis schemes for deep learning color image hiding algorithms.Therefore,this paper presents an image steganalysis algorithm based on a generative steganalysis rich model and Gabor residual network(SG-Res Net),which able to realize the classification and the preliminary extraction and reconstruction of hidden information.First,the carrier or stego images was fed into steganalysis rich model(SRM)to focus on the residual features of steganographic region and suppress the background content of the image itself.2D-Gabor filtering kernel convolution can learn many different types of dependencies between pixels in the residual image,so that the network can extract enough statistical attributes to accurately detect secret information.In this way,we able to minimize the negative effect of the high-pass filtering under the condition of guaranteeing the convergence of the network and output classification result.A reconstruction network also based on the SG-Res Net was designed to reconstruct the secret images from the stego image.Which includes the SRM convolution layer of channel separation and multiple pixel shuffle residual layers,and the hidden secret image was obtained through the successive extraction and reconstruction of these structures.The extensive experiments shown that the proposed SG-Res Net still has a good performance in terms of low hiding payload.(2)Generative Focused Feedback Residual Network for Image Steganalysis and Hidden Information Reconstruction.In response to the problem of high-resolution extraction and reconstruction of hidden information,this paper also proposed a generative image steganalysis algorithm based on focused feedback residual convolutional neural network for simultaneous detection and extraction of hidden information.First,a possible stego image was processed by a preprocessing network consisting of several convolutional layers and two fresh focus modules,and several enhanced feature maps were output.Then,the resulting enhanced feature maps were fetched into both a classification network and a reconstruction network to improve the sensitivity of implicit residual information in low-level semantic information.This makes it possible to capture multi-level features while increasing the receptive field of the steganographic region,and improves the detection accuracy.The classification network identified whether the feature maps came from a stego image or a simple cover image.The reconstruction network,consisting of some layers of convolutional units,pixel shufflers and feedback residual modules,completed the reconstruction of hidden information.Experimental results show that the proposed image steganalysis algorithm can obtain state-of-the-art results in terms of detection rate and hidden information reconstruction compared with classical rich models and several recent deep learningbased methods.