Digital Image Forensics Based On Homology Identification

With the advent of digital age, different kinds of multimedia communication ways greatly enrich the physical world and people’s real life. Digital image with its characteristics of strong visibility, small storage, easy to editing and modifying has been widely applied. With the rapid development of the Internet, the sensational effect caused by digital image is far greater than traditional media."Without image, without truth" has become a microcosm of the way people obtain information nowadays. With processing and editing software of digital image bringing convenience to people, it also greatly reduces the reliability of digital image. Therefore, digital image’s originality, authenticity and integrity authentication have become an urgent problem to be solved.Based on digital image passive forensics system, this paper proposed a framework of digital image forensics technique based on homology identification, which can implement complete image tamper detection and source identification via self-homology identification and cross-homology identification of digital images. This paper begins with the digital imaging process, then analyzes features introduced by the imaging process. Pattern noise (PRNU) is selected as the unique identification feature used for image homology identification. Aiming at the two aspects of self-homology identification and cross-homology identification, we proposed two forensics algorithm based on PRNU used for tamper detection and source identification. The main work of this paper and research results have been obtained are as follows:(1) Proposed a fast and unsupervised image source identification method based on graphFirstly, this paper analyzes the limitations of the existing source forensics algorithms and points out that most of the existing algorithms need prior knowledge for training classifier, which caused great limitation to forensics. Because of the above problems, this paper proposed a graph-based fast and unsupervised image clustering method. The algorithm uses spectral clustering algorithm to cluster affinity matrix of image set and gets the classification result. Some instabilities of this method are found in the experiment, so the optimal solution based on silhouette coefficient is proposed to improve it. The improved algorithm has good stability and veracity. Finally, contrast experiments are carried out to confirm the accuracy of the algorithm, the average accurate rate of the proposed method is about97%.(2) Proposed an improved image tamper detection method based on PRNUFirstly, this paper analyzes two problems of the existing algorithms:how to improve the correlation measure when image block is small and how to solve the contradiction in choosing the size of image block. To solve the two problems as said above, the peak correlation energy rate-PCE is proposed to use as the measure of correlation and an new image block dividing principle based on image saliency is proposed. Finally, experiments are carried out to confirm the accuracy of the algorithm.