Source Identification Forensics Of Natural Images And Computer Generated Graphics Based On PRNU

With the rapid development of digital devices and imaging software, it is very easy to acquire and modify digital images. Although these advanced technologies make people’s life more convenient, they also introduce a lot of security issues. If forged digital images are used in formal occasions such as news, evidence and research results by counterfeiters, serious implications on the authenticity of the events and the stability of society will be triggered. Hence, the study on how to ensure the authenticity and integrity of digital images is becoming a research hotspot.This thesis is focused on the research of identifying natural images (NI) and computer generated graphics (CG). Firstly, the background, significance and the state-of-art of digital image forensic technology are elaborated in the introduction section. What’s more, the main research contents and achievements are also reviewed. Secondly, related theory used in this thesis is briefly described. Finally, two novel methods are proposed.1. A novel scheme of identifying natural images and computer generated graphics based on hybrid features is proposed. Since the image acquisition pipelines are different, there exist some intrinsic differences in statistical, visual and noise characteristics between natural images and computer generated graphics. Firstly, the mean, variance, kurtosis, skew-ness and median of the histograms of grayscale image in the spatial and wavelet domain are selected as statistical features. Secondly, the fractal dimensions of grayscale image and wavelet sub-bands are extracted as visual features. Thirdly, considering the defects of the photo response non-uniformity noise (PRNU) acquired from wavelet based de-noising filter, a pre-processing of Gaussian high pass filter is applied to the image before the extraction of PRNU, and the physical features are calculated from the enhanced PRNU, with a total of48dimensions. In the identification, a support vector machine (SVM) classifier is used in experiments and an average classification accuracy of94.29%is achieved, where the classification accuracy for computer generated graphics is97.30%and that for natural images is91.28%. Analysis and discussion show that the method is suitable for the identification of natural images and computer generated graphics, and can achieve better identification accuracy than the existing methods with fewer dimensions of features. 2. A novel method of identifying natural images and computer generated graphics based on the characteristics of PRNU and the color filter array (CFA) interpolation is proposed. Firstly, based on the properties that CFA interpolation is a special operation of natural images and PRNU is a digital fingerprint of a camera, the differences of the impact of CFA interpolation on PRNU between these two types of images are analyzed. Secondly, the histogram of neighborhood variance of PRNU is calculated to represent them. Thirdly, in the RGB three color channels, the sum of the PRNU neighborhood variance and the max, the weighted average and the variance of its histogram are exacted respectively, with a total of12dimensions. Finally, a SVM classifier is used in experiments and an average classification accuracy of96.55%is achieved, which demonstrates that the proposed method is effective for the identification of natural images and computer generated graphics.These two algorithms proposed in this thesis can effectively and accurately identify natural images and computer generated graphics. Hence, they have a positive significance for the research of digital image passive blind forensics technology.