Compressed Multimedia Forgery Detection through Blind Digital Forensics

In today’s cyber world, digital images and videos act as the most frequently transmitted information carriers. However, the present day easy availability of low–cost image and video processing software and desktop tools, having immense number of multimedia manipulating features, pose threat to the fidelity of digital multimedia data. Digital images and videos being the major sources of evidence towards faithfulness of any event for law enforcement across the world, as well as in media and broadcast industries, maintenance of their trustworthiness and reliability is a major challenge in today’s digital world. This research deals with compressed image and video forgery detection. Almost all present–day digital devices use lossy/lossless compression formats for data storage. This research aims to exploit the effects of lossy/lossless compressions in multimedia files for detection of forgery. Joint Photographic Experts Group (JPEG) re-compression based image forensics has been a largely investigated area in the forensic community, over the recent years. This mainly involves investigating different compression quality factors in an image, by exploiting artifacts left behind by the compression stages. This thesis presents a deep Convolutional Neural Network model, that detects both double as well as triple compression in JPEG images. This thesis also addresses the major present-day challenges associated with video forgery detection, aimed towards solving the practical problems encountered by a forensic analysts during video tampering detection. Video forgery can be broadly classified into inter-frame and intra-frame forgery types based on their operational principles. This research focuses on the development and performance analysis of inter- as well as intra-frame video forgery detection and localization techniques, while addressing the certain additional challenges i.e., scenarios where number of forged frames is integral multiple of Group of Pictures (GOP), video compressed with dynamic GOP etc. This thesis presents statistical as well as deep learning based inter-frame forgery detection techniques. This thesis also investigates and develops a capsule network based forensic technique for detection of object based intra-frame forgery in surveillance videos, while improving the performance for low bit rate compressed videos