Image Statistical Frameworks for Digital Image

IMAGE STATISTICAL FRAMEWORKS FOR DIGITAL IMAGE FORENSICS by Patchara Sutthiwan The advances of digital cameras, scanners, printers, image editing tools, smartphones, tablet personal computers as well as high-speed networks have made a digital image a conventional medium for visual information. Creation, duplication, distribution, or tampering of such a medium can be easily done, which calls for the necessity to be able to trace back the authenticity or history of the medium. Digital image forensics is an emerging research area that aims to resolve the imposed problem and has grown in popularity over the past decade. On the other hand, anti-forensics has emerged over the past few years as a relatively new branch of research, aiming at revealing the weakness of the forensic technology. These two sides of research move digital image forensic technologies to the next higher level. Three major contributions are presented in this dissertation as follows. First, an effective multi-resolution image statistical framework for digital image forensics of passive-blind nature is presented in the frequency domain. The image statistical framework is generated by applying Markovian rake transform to image luminance component. Markovian rake transform is the applications of Markov process to difference arrays which are derived from the quantized block discrete cosine transform 2-D arrays with multiple block sizes. The efficacy and universality of the framework is then evaluated in two major applications of digital image forensics: 1) digital image tampering detection; 2) classification of computer graphics and photographic images. Second, a simple yet effective anti-forensic scheme is proposed, capable of obfuscating double JPEG compression artifacts, which may vital information for image forensics, for instance, digital image tampering detection. Shrink-and-zoom (SAZ) attack, the proposed scheme, is simply based on image resizing and bilinear interpolation. The effectiveness of SAZ has been evaluated over two promising double JPEG compression schemes and the outcome reveals that the proposed scheme is effective, especially in the cases that the first quality factor is lower than the second quality factor. Third, an advanced textural image statistical framework in the spatial domain is proposed, utilizing local binary pattern (LBP) schemes to model local image statistics on various kinds of residual images including higher-order ones. The proposed framework can be implemented either in singleor multi-resolution setting depending on the nature of application of interest. The efficacy of the proposed framework is evaluated on two forensic applications: 1) steganalysis with emphasis on HUGO (Highly Undetectable Steganography), an advanced steganographic scheme embedding hidden data in a content-adaptive manner locally into some image regions which are difficult for modeling image statics; 2) image recapture detection (IRD). The outcomes of the evaluations suggest that the proposed framework is effective, not only for detecting local changes which is in line with the nature of HUGO, but also for detecting global difference (the nature of IRD). IMAGE STATISTICAL FRAMEWORKS FOR DIGITAL IMAGE FORENSICS by Patchara Sutthiwan A Dissertation Submitted to the Faculty of New Jersey Institute of Technology in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy in Electrical Engineering Department of Electrical and Computer Engineering