Image Steganalysis via Transfer Learning Framework

Abstract

In the real world, millions of images are captured by cameras and mobile phones that have cameras, and different users are transferring these images daily in the virtual world. Due to the extensive advancement of science and technology, data security has become very important. One of the main methods of maintaining data security is steganography. The concept of steganograohy is to hide information, in a cover media such as an image. Hiding information may be used by profiteers and misused. For this reason, the issue of steganalisys, which seeks to find this hidden information in the image, is raised. In recent years, various steganalisys algorithms have been proposed to ensure the security of transmitted information in the Internet space to which everyone has access, and have achieved good results. The steganalysis on Internet images will encounter steganographic algorithm mismatch (SAM) and cover source mismatch (CSM). Therefore, the steganalysis on the Internet is essentially to solve the mis- match problem. The problem can be solved according to the different domains in which the images are placed, using transfer learning methods. Training a classifier, by using training data and labeling testing data, despite the difference in distribution between training and testing data, reduces the efficiency of machine learning algorithms. Transfer learning and domain adaptation is one way to solve the problem of domains transfer, in which there is a large distributional difference between the training data or the source domain and the testing data or the target domain. The purpose of visual domain adaptation is to learn robust models for the target domain, using knowledge transfer from the source domain. Therefore, domain adaptation increases the efficiency of the trained model across the source domain in order to label target domain samples. This method facilitates our work in the problem of steganalisys and by using this method the problem of mismatch can be solved. In this thesis, a method based on unsupervised domain adaptation is proposed. This makes the distribution between the training and testing sets more similar to achieve better diagnostic performance. We use a standard classification model and create a separate subspace and predict the model error to match the domains and predict the label of the target domain samples. To match the distribution between domains we used the MMD method, which includes marginal and conditional distributions. The results obtained after implementing the method in experiments (including SAM and CSM) on the designed database, show that our method performs better than the existing advanced methods.

Keywords: Steganography, Steganalisys, Transfer Learning, Domain Adaptation, Distribution Adaptation, Subspace Adaptation, Image Classification.