Research On Image Statistical Modeling And Its Application To Image Forensics

Image forensics, as a significant part of information security, is a technique which aims at identifying the forgery, tampering and steganography by analyzing the statistical features of digital images. Many problems in this field can be assumed to be a statistical pattern recognition issue and therefore statistical modeling of images plays a very important role in image forensics. In this paper, the statistical modeling method based on generalized Gaussian distribution is discussed and then the discrimination between natural images and photorealistic computer graphics, LSB matching steganalysis and the impact of image complexity on steganalysis performance are studied. The main work and contributions of this thesis are summarized as follows:1. A new discrimination method using second-order difference statistics is proposed to distinguish natural images from photorealistic computer graphics. Firstly, the second-order difference signals and predicting error signals of both original and calibrated images are derived in the HSV color space, and then the variance and kurtosis of the second-order difference signals and the first four order statistics of the predicting error signals are extracted to be used as distinguishing features, the Fisher linear discriminant is applied to constructing a classifier to do the differentiating job. Experimental results show that the proposed method exhibits excellent performance for the discrimination between natural images and computer graphics. Moreover, it has a low computational complexity.2. A LSB matching steganalysis method is proposed on the basis of statistical modeling of pixel difference distributions. As is indicated in the former researches, natural images are highly correlated in a local neighborhood, the value zero appears most frequently in intensity differences between adjacent pixels, and the statistical model of the distribution of pixel difference can be established using the Laplace distribution. As LSB matching steganography randomly increases or decreases the pixel value by 1 when the message is embedded, the frequency of occurrence of the value zero in pixel differences changes most dramatically during message embedding. Based on the Laplace model of pixel difference distributions, this thesis proposes a method to estimate the number of the zero difference value using the number of non-zero difference values from stego-images, and uses the relative estimation error between the estimated and actual values of the number of the zero difference value as the classification characteristic. Experimental results indicate that the proposed algorithm is effective to detect LSB matching steganography and can achieve better detection performance than the local extreme method under most circumstances.3. The impact of the image complexity on steganalysis performance is discussed. The generalized Guassian distribution model is used to model the pixel difference distributions and the shape parameter is defined to describe the image complexity. The Bhattacharyya distance is applied to measure the steganalysis performance of a classification feature. Then the impact of image complexity on feature value and Bhattacharyya distance is analyzed. Both the theoretic analysis and experimental results show that the smaller the image complexity is, the better the steganalysis performance of classification characteristic will be.Finally, the research work in this thesis is summarized and the further research topics and directions in the future are prospected.