Research On Blind Digital Image Steganalysis Based On Feature

Steganalysis is an important topic of hiding communication in information security area, which is widely used in military and national security. Digital image is one of the most common cover medium in the digital information, and the development of digital image steganography has been mature. By contrast, research in digital image steganalysis has got some research achievements, but there are still some way from practical application, especially the blind steganalysis for the digital image. Due to the quick update and the abundant methods of the digital image steganography, makes blind steganalysis for the digital image become a challenging research topic.This paper concentrates on the universal blind detection in digital image steganography. Firstly, the basic framework of the steganographic blind detection of digital image was introduced, and a theoretical analysis and summary about the feature used in the typical steganographic blind detection method were made. Then,we focus on the analysis of the principle of feature calibration to improve the performance of features, and conclude the reason of discontinuous phenomenon in the stego images from the basic idea of the crop calibration and the prediction error calibration. On this basis, A method of feature calibration based on sub-pixel translation is proposed. The method uses sub-pixel translation to remove the existence information of steganography in images, and improves the discrimination of features. Experiments show that the proposed method can effectively improve the performance of feature, and then improve the accuracy of detection algorithms. Furthermore, this paper also proposed a universal blind steganalytic algorithm based on multi-domain feature by combining multi-domain feature with feature calibration The algorithm uses multi-domain feature to represent the effect of different steganography in images, and introduces the sub-pixel calibration proposed by this paper to improve the discrimination of features, and then the support vector machine (SVM) was adopted as the classifier. A large number of experiments in binary classification of single algorithm, binary classification of mixed algorithm and ten classification of mixed algorithm show that the proposed algorithm get a good binary classification results, and has a certain ability to distinguish as many as ten classes of common steganography approaches.