Image Steganalysis Based On Deep Convolutional Neural Network

Digital images are the major covers for modern steganography.The huge amount of image data and a wide range of image steganography tools on the Internet make it easier for people to embed secret information in images.Image steganography can be used not only to protect legitimate activities such as private communications,but also to cover up illegal activities such as commercial espionage and transmitting terrorist attacks.If image steganography is used maliciously,it will cause extremely serious social harm.However,complex application scenarios and huge amounts of data have presented great challenge to image steganalysis.The actual steganalysis needs to detect a large number of JPEG images,but most of the existing JPEG steganalysis methods tested on small-scale and single source image data sets.It is difficult to ensure that there is no overfitting problem.Besides,researchers currently pay less attention to color images steganalysis,while most of the image data in the real world are color images.Therefore,it is of great significance to study JPEG image steganalysis and color image steganalysis suitable for big data scenarios.With the increase of data,deep learning can obtain better classification results and model generalization ability,which is very suitable for the practical application of image steganalysis in the context of big data.Motivated by above analysis,this paper conducts research on image steganalysis based on convolutional neural networks.The main contributions in this dissertation are as follows:(1)To have proposed a deep learning-based pre-training network to simulate the extraction procedure of rich-model features for JPEG image steganalysis.At present,the domain knowledge behind steganalysis cannot be used more effectively in deep learning.Considering that the rich model JPEG image steganalysis has certain similarities with convolutional neural networks,we propose a pre-trained subnetwork via fitting deep neural network to rich-model features extraction procedure.We select DCTR(Discrete Cosine Transform Residual)feature set with good performance and low computational complexity,as the fitting object.The DCTR feature extraction process includes multiple sub-modules.Based on these sub-modules,we propose deep learning sub-networks with similar functions and pre-train these sub-networks.After pre-training,the proposed network can obtain better steganalysis performance compare to the DCTR feature.The experimental results show that the proposed pre-training network is able to fit the nonlinear mapping implicit in the DCTR feature extraction process.When continues training and finetune the network from a pre-trained state similar to DCTR feature set,the proposed network can obtain far better detection performance than DCTR feature.(2)Large-scale JPEG image steganalysis using hybrid deep learning framework.The main difficulty of JPEG images steganalysis lies in the need to detect large-scale JPEG image data with complex sources.However,the detection performance of existing JPEG image steganalysis rich model feature will significantly drop when it comes to large-scale JPEG image data with complex sources.We adopt a more effective strategies to combine deep learning and JPEG image steganalysis domain knowledge,and propose a hybrid deep learning framework for large-scale JPEG image steganalysis.The hybrid framework consists of two main stages.The first stage introduces a frequency domain filtering kernel convolutional layer and a diverse quantization & truncation layer.The second stage is a compound deep-neural network containing multiple deep subnets.We provide experimental and theoretical proofs to support the introduction of structures such as frequency domain filtering kernel convolution layers and quantization & truncation layers.In experiment part,we constructed a large-scale JPEG image set containing 5 million JPEG images to verify the performance of the whole framework in a big data environment.The experimental results show that the proposed framework is superior to all existing image steganalysis algorithms,and can adapt to the image steganalysis problems in complex application scenarios such as continuously expanding data scale and image source mismatch.At the same time,the proposed framework is extensible.For example,the newly emerging deep learning steganalysis network can be incorporated into the proposed framework as a subnet configuration.(3)WISERNet: Wider Separate-Then-Reunion Network for steganalysis of color images.At present,there is no solution for color images steganalysis based on deep learning.In addition to applying deep learning to JPEG image steganalysis,this paper also proposes a deep learning solution for color image steganalysis for the first time.We propose a Wider Separate-Then-Reunion Network,which can be abbreviated as WISERNet.We provide theoretical rationale to claim that the summation in normal convolution is one sort of linear collusion attack which reserves strong correlated patterns while impairs uncorrelated noises.Based on this,we adopt separate channel-wise convolution in the bottom layer as the first ”separation” part of WISERNet.In the upper convolutional layer,a wider conventional convolution structure with more convolution kernels is implemented to enhance the classification capability.Which can be regarded ad the post ”aggregation” part of WISERNet.To fully verify the practical performance of WISERNet,we constructed a variety of color image datasets generated using different demosaic and downsampling algorithms in the experiment part.Experimental results show that WISERNet is particularly suitable for color image steganalysis,which has huge advantages compare to other existing color image steganalysis algorithms.This paper had made some progress in image steganalysis.Looking into the future,there are still many problems in image steganalysis that need further study.It is our mission to continue develop practical steganalysis to prevent image steganography from being maliciously used.