Research On Blind Detection Based On Image Content

As an art of covert communication, steganography tires to transfer secret messages througha public channel by hiding the secret data into an innocuous-looking cover medium such as adigital image. In contrast to steganography, steganalysis aims to supervise or destroy the covertcommunication by detecting the presence of the secret messages through statistical inspection.As an important category of steganalysis, blind detection, which can be seen as a patternrecognition problem, has much flexibility and can be adapted to new steganography algorithms.Such a research field has drawn extensive attention in the multimedia security community．Digital images, carrying the visual information of different scenes with specific spacialstructures, have local stationary characteristics. Because the smoother the image content is, themore obvious the statistical differences are. Steganalysis mainly depends on a change ofstatistical characteristics bwteen cover images and stego images, and therefore the steganalyticfeatures are closely related to the image content. Consequently, it is of great significance toconduct research on the image universal steganalysis combined with image content.On the basis of the analysis of the relationship between the steganalytic feature and theimage content, the thesis focuses on the study of blind detection based on image content. Thecontributions obtained in this thesis can be summarized as follows:1. The relationship between the complexity of the image content and the statistical feature isanalyzed by employing information theory and statistics theory. The analytic results demonstratethat images have strong spatial dependences among adjacent pixels and images with differentcontent present different statistical characteristics.2. On the self-built database with single image content, the influences of LSB matching andF5steganography on the statistical characteristics of images are analyzed. Then, the detailcomponents are modeled as the General Guassian Distribution in order to investigate therelationship between the complexity of image content and the steganalytic features especially forthe blind steganalytic features. The analyses show that the steganalytic features have a closerelationship with the image content, which provides a theoretical basis for studying the imagecontent based universal blind detection techniques and proposing new reliable steganalyticalgorithms.3. In allusion to the current steganographic methods for JPEG images, the distribution andstatistical features as well as intrablock and interblock correlations of Discrete Cosine Transform(DCT) coefficients are analyzed. In succession we explore the extraction and selection offeatures to further propose a blind steganalysis approach for JPEG images based on regioncorrelation. First, the macroscopic and microscopic calibrations are combined to estimate coverimages. Then, intrablock and interblock correlations of DCT coefficients are modeled as Markov.Finally, the differences of transition probability matrices are used as features for a steganalyzerimplemented by support vector machines, realizing the detection of current steganographicalgorithms for JPEG format. The experimental results on joint database present that the proposedsteganalyzer outperforms the existing typical steganalyzers. 4. From the viewpoint of time (space)-frequency analysis of image source, a universalsteganalytic scheme is proposed based on Contourlet transform, which could stands for multi-scale and localization of the image. Firstly, the difference between the characteristic function(CF) of detail components before and after embedding is analyzed in order to construct a kind ofnovel band pass CF moment. Then, the CF moments of image and prediction errors are extractedby comparing and analyzing influences of two types of image prediction methods on detectionpower. Finally, we utilize SVM to classify cover and stego images. Extensive experimentalresults on BOWS and UCID databases show that the proposed method outperforms the existingtypical steganalysis methods.5. Considering the influences of the complexity of image content on the steganalyticfeatures, a universal feature fusion steganalytic scheme is proposed based on the complexity ofimage content. First, different steganalytic features are extracted according to the imagecomplexity. Then, Bhattacharyya Distance is used to evaluate the usefulness of features in orderto obtain the weights of individual feature. Consequently, the weights got above are fusioned todetect the current steganographic algorithms for JPEG format. Experimental results on jointdatabase demonstrate that the proposed steganalyzer not only reduce computational complexity,but also provides a more reliable detection.Finally, the research work for this thesis is summarized and the further research topics anddirections in the future of blind detection of information hiding are discussed.