Image Steganalysis In Spatial Domain Based On Convolutional Neural Network

Image steganography is the technology of embedding secret messages into the elements of an image to make these messages undetectable.On the contrary,image steganalysis is aiming to detect the presence of secret messages in an image,and this is usually achieved through two steps,feature extraction and training classifier.Recently,as content-adaptive steganographic schemes become popular,secret messages are able to embedded into regions with complex content,which makes the embedding traces less detectable.To detect these traces,the features used in image steganalysis become inevitably more and more complicate and high-dimensional.And it is very challenging for researchers to extract rich and effective features.Compared to handcrafted feature based methods,Convolutional Neural Network(CNN)is able to learn effective features from training data automatically,without heavily relying on the domain knowledge of researchers.Moreover,the steps of feature extraction and training classifier could be optimized simultaneously easily when using CNN.To take the advantage of the powerful ability of CNN of learning and representing features,CNN-based image steganalysis methods are studied in this paper and the main works are listed below:1.A simple CNN architecture is proposed.Mainly constructed using two basic modules,a small convolutional block and a pooling layer,this architecture can be easily adjusted and expanded.This architecture could obtain good detection results when employing some useful practices proposed in other concern works.2.In this paper,it is proposed to use CNN to estimate the knowledge of selection channel,i.e.,the probability of each pixel being modified.Moreover,a multi-task learning structure is also proposed to further boost the performance of image steganalysis.This structure consists of specified branch for image steganalysis,specified branch for estimating the knowledge of selection channel and a backbone shared by them.By simultaneously learning the main task,image steganalysis,and the auxiliary task,estimating the knowledge of selection channel,the main task could be aware of the useful information contained in the training signal of the auxiliary task and obtain better generalization ability.To verify the effectiveness of the proposed methods,extensive experiments are conducted on three excellent steganographic algorithms and six payloads.And the experimental results show that the both of the two CNN-based methods proposed in this paper obtain lower detection error rate the maxSRM,the state of the art handcrafted feature based approach.