Abstract:

Digital images have become an important carrier for steganography, which is the science and art of covert communications. Hidden data makes a stego image deviate from its associated cover image statistically. Discovering the statistical difference between these two kinds of images is hence the key issue to image steganalysis, which is the science and art to detect the very existence of hidden data in digital images. In this research, image statistical models have been proposed and addressed with the framework of machine learning to fulfill the task of image steganalysis and some other tasks in image forensics.

Texture images are known of noisy nature and data hidden in raw texture images is hence hard to detect. Using statistical moments of characteristic functions generated from both one- and two-dimensional histograms and what is called rake transform has resulted in an effective texture image seganalyzer, which has outperformed the existing steganalyzers designed for steganalyzing texture and smooth images.

Because the interchange of JPEG images is most popular nowadays, JPEG image steganalysis is of significance. Three JPEG image steganalyzers have been proposed. Applying the above mentioned statistical moments to both image spatial and JPEG domains has improved JPEG steganalytic capability remarkably. A more effective JPEG steganalyzer uses Markov process applied to the difference JPEG 2-D arrays, i.e., the differences between the magnitude array of the JPEG coefficients and its shifted versions along different directions, and outperforms the state of the art. Utilizing both intrablock and interblock correlation among JPEG coefficients has proved to be able to further boost JPEG image steganalysis performance.

The above statistical models have shed light to other image forensic tasks. An approach based on difference JPEG 2-D array and Markov process to identifying double JPEG compressed images and an approach combining statistical moment and Markov process features derived from both the spatial and rake transform domains to image splicing detection are shown to have the performance exceeding the prior arts.

Committee Members:

Dr. Yun Q. Shi, Professor, Department of Electrical and Computer Engineering, NJIT (Advisor)
Dr. Jeffrey A. Bloom, Manager, Content Security Research Group, Thomson Corporate Research
Dr. Hongya Ge, Associate Professor, Department of Electrical and Computer Engineering, NJIT
Dr. Richard A. Haddad, Professor, Department of Electrical and Computer Engineering, NJIT
Dr. Frank Y. Shih, Professor, Computer Science Department, NJIT
Dr. Edward K. Wong, Associate Professor, Department of Computer and Information Science, Polytechnic Institute of New York University

Click here for seminar archive

Note: All MS thesis defense and PhD dissertation (proposal) defense are counted towards ECE791.