Research On Multi-view Features Of Image Steganalysis Technique

With the rapid development of computer science and web multimedia technology, Information Security gets more and more attention. Steganalysis is a new branch of information security, which mainly researches on the effective detection of the cover image and boosts the development of information security. Especially after the 9.11 event, steganalysis became more important, which plays an important role in national defence and military security. From the former LSB-based method to the universal algorithms, steganalysis has got great development in recent years, and provides maturely schemes for the information detection. Therefore, research on steganalysis is very valuable both in academic development and potential applications.In this dissertation, we firstly introduce the basic theory and the development of information hiding and information security. The concept and the theory models of steganography and steganalysis are presented subsequently. Then we focus on the state-of-art of steganalysis and point out the current problems. Based on the above study, we mainly research on blind/universal steganalysis technique, and achieve some valuable results as follows:(1) By analyzing the changes of the features between cover and stego images from DCT domain, DWT domain, DFT domain and spatial domain, a series of features extracting algrothimies are exploited depending on DCT domain, DWT domain, DFT domain and spation domain respectively. We firstly introduce image prediction technique, and then extract features'cofficients from predicted images using the same algrothimies as we proposed before. In the end, these features become another new vector which can be fed to a SVM classifier.(2) An ONPP based random subspace ensemble strategy is proposed in this thesis. Because of the high dimension of extracted features in this algrothimy, datas with high dimension usually easily cause a big disaster in classifier. The ONPP is introduced in this paper which is a linear dimensionality reduction technique, which will tend to preserve not only the locality but also the global geometry of the highdimensional data samples. It can be extended to a supervised method, and it can also be combined with kernel techniques. This reducing dimension algrothimy were introduced which can decrease the complex of the whole system and save more time to process amount of data.(3) Design the classifier. Due to the principle and the performance of the classifier, and the effectiveness of the feature vector. SVM is utilized as the classifier. Compared with other classifiers, such as "Nearest Neighbor" and then analysis the detecting results. The performance is very well.