Blind Forensics Based On Statistical Features Of Digital Image

There are many ways to get information for human, such as hearing, sight and taste. The information provided by eyes accounts for about 80%. Digital image, as the most common carrier of visual information, can obtain more accurate and true information than that obtained by other ways. However, with the popularity of digital cameras and personal computers, people get the pictures which could have been forged and tampered maliciously by image editing software. The important way to get reliable and authentic information is seriously threatened. Blind digital image forensics is a technology that can identify the authenticity and source of digital images. These images don't contain digital watermarking, digital signature or digital fingerprint. In this paper, we study the blind digital image based on statistical characteristic of image. We also propose a new algorithm to distinguish natural images from computer generated ones.Firstly, the influence of the selection of discrete cosine transform(DCT) coefficients on distinguishing natural images from computer generated ones is studied. We study it by the Benford's law. Simulation results show that for the same criteria and the optimal threshold, classification accuracy rate of two kinds of images firstly increases with the quantity of high-frequency DCT coefficients, then reaches its maximum, at last decreases. These results show that high-frequency DCT coefficients are more beneficial to improve classification accuracy rate of two kinds of images. For the above mentioned, accuracy rate of classification is the highest for gray image. It is less for the color G channel and worst for R and B channel.Secondly, the effects of the selection of the DCT coefficients on distinguishing natural images from spliced ones are studied. Simulation results show that selecting the DCT coefficients from the highest frequency to low frequency, with the low frequency components increased, accuracy rate of image classification changes relatively smaller. When selecting the DCT coefficients from the lowest frequency to high frequency, the accuracy rate of image classification is greatly improved. Therefore, the quality of high frequency components captures the feature of sharp edges of the spliced images. When choosing the same number of the DCT coefficients, the scheme for selecting DCT coefficients from the highest to low frequency is performed generally better than that for selecting DCT coefficients from the lowest to high frequency.Finally, a blind forensic algorithm for distinguish natural images from computer generated ones is proposed. The algorithm is based on the disorder of image hierarchical model. We extract the degree of disorder of each binary. These features of disorder are used to train a support vector machine(SVM), which can predict the category of digital images. Simulation results show that by dividing binary layer into the appropriate subsequence and computing disorder degree, SVM can use these features to distinguish natural images from computer generated ones. It has a classification accuracy rate of 82.80%. The area under the receiver operating characteristic curve(AUC) is 0.8982. The performance of classifier is good and it's just below the excellent. A good classification result is obtained by running this algorithm, which is used for distinguishing natural images from computer generated ones.