Multimedia Forensic Analysis via Intrinsic and Extrinsic Fingerprints

Abstract

Digital imaging has experienced tremendous growth in recent decades, and digital images have been used in a growing number of applications. With such increasing popularity of imaging devices and the availability of low-cost image editing software, the integrity of image content can no longer be taken for granted. A number of forensic and provenance questions often arise, including how an image was generated; from where an image was from; what has been done on the image since its creation, by whom, when and how. This thesis presents two different sets of techniques to address the problem via intrinsic and extrinsic fingerprints.

The first part of this thesis introduces a new methodology based on intrinsic fingerprints for forensic analysis of digital images. The proposed method is motivated by the observation that many processing operations, both inside and outside acquisition devices, leave distinct intrinsic traces on the final output data. We present methods to identify these intrinsic fingerprints via component forensic analysis, and demonstrate that these traces can serve as useful features for such forensic applications as to build a robust device identifier and to identify potential technology infringement or licensing.

Building upon component forensics, we develop a general authentication and provenance framework to reconstruct the processing history of digital images. We model post-device processing as a manipulation filter and estimate its coefficients using a linear time invariant approximation. Absence of in-device fingerprints, presence of new post-device fingerprints, or any inconsistencies in the estimated fingerprints across different regions of the test image all suggest that the image is not a direct device output and has possibly undergone some kind of processing, such as content tampering or steganographic embedding, after device capture.

While component forensics is widely applicable in a number of scenarios, it has performance limitations. To understand the fundamental limits of component forensics, we develop a new theoretical framework based on estimation and pattern classification theories, and define formal notions of forensic identifiability and classifiability of components. We show that the proposed framework provides a solid foundation to study information forensics and helps design optimal input patterns to improve parameter estimation accuracy via semi non-intrusive forensics.

The final part of the thesis investigates a complementing extrinsic approach via image hashing that can be used for content-based image authentication and other media security applications. We show that the proposed hashing algorithm is robust to common signal processing operations and present a systematic evaluation of the security of image hash against estimation and forgery attacks.