The Research On JPEG Image Mismatched Steganalysis Based On Feature Analysis

Steganography, mainly researches on how to embed secret information to be transferred or hidden into digital media such as images, audio, video and text, without causing other people’s suspicion. Steganalysis is a countermeasure technology to steganography and aims to detect the presence or absence of hidden data in given media objects. With the development of the information age, when people benefit from steganography in privacy preserving, criminals also use it for malicious purpose. Therefore, the research on steganalysis is of great urgency. JPEG images are widely used in the field of steganography and steganalysis for the inherent excellent properties. The conventional JPEG steganalysis has achieved a good detection performance. But this success relies on the assumption that both training samples and testing samples are from the same source with similar statistical properties and feature distribution. However, this assumption is usually not true in practical steganalysis applications, which may result in the problem of mismatch and the degradation in detection accuracy.At present, there is little research on the JPEG mismatch steganalysis. In this paper, the main framework of the conventional JPEG steganalysis is introduced in detail, including the image database, steganalysis feature and steganalysis classfier. Based on on the previous research, we explain the definition of the mismatch, the cause of mismatch and the different types of mismatch, and present the currently existing methods for mitigating the mismatch problem. In addition, we verify the existence of the mismatch problem by taking the quantization table mismatch as an example. In this paper, we study the problem from the view of feature analysis, and the contribution of this paper is summarized as follows:1.In chapter 3, we propose three complementary criterions for JPEG mismatch steganalysis as follows:improving the statistical consistency, reducing the difference of feature distribution and preserving the classification ability of the training data. Based on the criterions, we put forward a novel mismatched steganalysis algorithm, Robust Discriminative Feature Transformation. It deals with the steganalysis feature from three aspects, feature alignment, minimizing the feature discrepancy and maximizing the feature discriminability of training data, which can help to mitigate the problem of mismatch. In addition, we design and conduct experiments on three kinds of mismatched scenarios, quantization table mismatch, steganographic algorithm mismatch and embedding rate mismatch. And Experimental results illustrate that the proposed method outperforms the prior arts and its detection accuracy of our proposed method is close to that of matched scenario. We also conduct experiments on small sample quantity and parameter robust.2. In chapter 4, we propose another JPEG mismatch steganalysis algorithms from the view of feature matching:domain invariant feature learning. It can mitigate the mismatch problem by the domain invariant feature which can be obtained by minimizing the square loss between the estimated kernel feature matrix of the testing data and the kernel feature matrix of the training data. We design and conduct experiments on three kinds of mismatched scenarios. And experimental results illustrate that these two proposed methods can mitigate the mismatch problem and outperform the prior arts. We also conduct experiments on multi-factor mismatch, which indicate they can mitigate the multi-factor mismatch.3. In chapter 5, we propose another JPEG mismatch steganalysis algorithms:jointly fusing feature matching and sample selection. It reduces the mismatch between training and testing data by jointly uniting the feature matching and sample selection in the kernel-PCA. We design and conduct experiments on three kinds of mismatched scenarios. And experimental results illustrate that these two proposed methods can mitigate the mismatch problem and outperform the prior arts. We also conduct experiments on multi-factor mismatch and compare these three methods in the paper.