Research On Jpeg Image Based Steganographic Codes And Blind Steganalysis

With the rapid development of the network technology, communication technology and multimedia signal processing, as a new technique, information hiding becomes a cutting-edge of the information security area. As an important branch of information hiding, steganography aims at sending secret message covertly by embedding it into public digital multimedia. And steganographic code is the core of the steganography, which would embed as many messages as possible per change of the cover object. On the other hand, steganalysis studies on how to detect, extract, and recover/destroy the hidden messages.Targeting on development and application requirements of the information hiding technology, some researches are done on JPEG image based steganographic code and blind steganalysis in the present thesis. The bodies of the studies would be outlined as follows:1,Constructing MME+1 (Modified Matrix Encoding) steganographic code and two kinds of steganography based on MME+1.(1) The proposed MME+1 code combines different MME codes to embed one more bit into the cover of the same length. Without bring more computational complexity, MME+1 has higher embedding efficiency than MME.(2) Based on (1), LSB Matching (Least Significant Bit Matching) steganography based on MME+1 is presented for JPEG decompressed image. Experiments show that the algorithm reduces the rounded disorder of the cover coefficients, and enhances the capability resisting the detection method targeting on LSB Matching.(3) Based on (1), IRB steganography based on MME+1 is presented on frequency domain. Experiments show that the algorithm reduces the rounded disorder of the cover DCT coefficients, and enhances the capability resisting blind detections based on calibration.2,Constructing blind steganalysis based on multi-directional correlations of the DCT (Discrete Cosine Transform) coefficients and calibrations.(1) A novel kind of feature matrix is constructed to describe correlations of the DCT coeffi- cients in multi-directions. Then, by merging predicted error calibration and image cropped calibration, a 96-dimensional feature vector is extracted. A series of experiments are performed on 6 kinds of typical steganography: F5,OutGuess(ver 2.0),JpHide(ver 0.5),StegHide,MB1 and MB2, showing that,the 96-dimensional feature vector is sensitive to those steganography algorithms.(2) The 96-dimensional feature vector is sent to the SVM(Support Vector Machine) to build the two-class steganalyzer to classify cover and stego images. Experiments show that the steganalyzer, which is against F5,OutGuess(ver 2.0),JpHide(ver 0.5),StegHide, MB1 and MB2, would classify the cover and stego with high accuracy.(3) The 96-dimensional feature vector is sent to the SVM(Support Vector Machine) to build the multi-class steganalyzer to classify JPEG steganography algorithms. Experiments show that the steganalyzer would classify those algorithms in some extent. The proposed method laies the foundation for the future extracting the secret messages.