Steganalysis Research Based On Image Region Statistical Features

As one of the most important technical tools for defending the security of imageinformation, image steganalysis has evolved into an attractive research topic in the field ofmultimedia information security. Distinguished from the existing steganalyzers based on theentire image data, the thesis focuses on analyzing the region statistic features of image contentand the relationship between the steganalytic features and image content so that imagesteganalysis algorithms are developed around the image region statistic properties by utilizingthe techniques for randomness measurement, textural description and image segmentation in thisdissertation. The main contributions of this thesis are summarized as follows:1. Analyses of several statistical features of natural images and the relationship betweenthe existence of the hidden messages and these statistical features. First of all, on the basis of theregion stationary Markov model of natural images and three self-built image databases withsingle content, the differences of statistical features of classified images are analyzed by adoptingthe relative techniques for randomness testing, probability and statistics theory, informationtheory and image modeling. These statistical features include the randomness features of thebitplane, the picture information measurement features and spectrum features of the gray pixelhistogram, the statistical modeling features of the pixel difference histogram, and grayco-occurrence matrix features. Moreover, after constructing the stego images by applying LSB(Least Significant Bit) matching steganography to self-built classified images, we analyzetheoretically and demonstrate experimentally the differences of some typical steganalyticfeatures between the cover image and its counterpart the stego image. Extensive experimentalresults conclude that steganalytic features are not only influenced by the embedding operation,but also are closely related to the image content. The smoother the image content is, the moreevident the statistical differences are. Last but not the least, based on these analytic conclusionsabove, a scheme for exploring the steganalysis technique by employing image regional content isachieved, and the guidelines and technical lines that support the concrete research work areilluminated.2. A detection approach of LSB matching based on image regional randomnessmeasurement. Since the image source possesses the region stationary characteristic, its regionpixel values have similar statistical relation which can be utilized to measure the imagerandomness in order to improve the sensitivity of steganalytic features. By deep analyzing theinfluences of the LSB matching embedding on the images with different contents, an effectivesteganalysis method is proposed in accordance with the features of image regional randomnessmeasurement. The concrete operations include separating an image into blocks, pixel Hilbertscanning and LSB sequence bit xor for defining the index of region randomness measurement,extracting four sorts of features of the index probability distribution for the classification of theFisher linear discrimination. Experimental results demonstrate that the detection performance ofthe proposed approach is improved compared with the four congeneric approaches.3. A blind steganalytic algorithm for spatial domain steganography based on region textural features. Based on the effective description of image region content by texture, theimage texture analysis can be applied to the steganalysis, and the stego noise can be regarded asa kind of additive stochastic texture in a fine scale. From the image texture viewpoint, asteganalytic method to detect spatial domain steganography in grayscale images is proposeddepending on the relationship between image region texture features and steganalytic features. Inthe first place, the local linear vectors are constructed by selecting carefully the Laws texturaldescription vectors and discrete cosine transform basis vectors. Then the image is decomposedinto a group of textural description subbands by the local linear transform (LLT) in fourdirections, and the statistical distributions of the LLT coefficients are modeled by using thegeneralized Gaussian distribution (GGD). Finally, novel texture features of the histogram and theco-occurrence matrix of LLT coefficients are extracted and then their dimensions are reduceddepending on the theoretical analysis of GGD model of LLT coefficients. A steganalyzer isimplemented by combining the texture features and the support vector machine classifier.Extensive experiments are performed on four diverse uncompressed image databases and seventypical spatial domain steganographic algorithms. It reveals that the proposed scheme isuniversal in detecting spatial domain steganography. In comparison with other three well-knownfeature sets, our presented feature set offers the best performance under most circumstances.4. A research on the synthetical discrimination steganalysis scheme based on the imageregion segmentation. The different sub-regions of the image exhibit diverse statistical characters,while the secret messages involved in the steganography possess the statistical consistency.Therefore, the identical steganalytic feature may present different changes in different regions,and different steganalytic features may exhibit different detection performances in the sameregion. Based on analyzing the connection between the existence of hidden messages and imageregion content statistical features, a new steganalysis scheme is designed on the basis of theimage region statistical characters and image segmentation technique. The main principle of thescheme is to divide an image into several categories of sub-regions, analyze the effects of thesteganography on each category of sub-regions respectively, and achieve a novel steganalyzerbased on the synthetical discrimination. Under the guidance of the proposed scheme, a concretedetection algorithm is realized against the LSB matching steganography, where the imagecontent complexity is defined for image segmentation, the cover and stego spectrums of eachkind of sub-region are analyzed, spectrum features are extracted by bandpass filters, and Bayesclassifiers are used to train and test. Experimental results demonstrate that the proposed methodexhibits good performance for the detection of LSB matching, and outperforms the existing ninerepresentative approaches.Finally, the research work for this thesis is summarized, the common relationships anddifferences among the four research points above are interpreted and the further research topicsof information hiding are discussed in the end.