Photo response non-uniformity based image forensics in the presence of challenging factors

Preview

PDF WRAP_Theses_Quan_2020.pdf - Submitted Version - Requires a PDF viewer. Download (3MB) | Preview

Request Changes to record.

Abstract

With the ever-increasing prevalence of digital imaging devices and the rapid development of networks, the sharing of digital images becomes ubiquitous in our daily life. However, the pervasiveness of powerful image-editing tools also makes the digital images an easy target for malicious manipulations. Thus, to prevent people from falling victims to fake information and trace the criminal activities, digital image forensics methods like source camera identification, source oriented image clustering and image forgery detections have been developed.

Photo response non-uniformity (PRNU), which is an intrinsic sensor noise arises due to the pixels non-uniform response to the incident, has been used as a powerful tool for image device fingerprinting. The forensic community has developed a vast number of PRNU-based methods in different fields of digital image forensics. However, with the technology advancement in digital photography, the emergence of photo-sharing social networking sites, as well as the anti-forensics attacks targeting the PRNU, it brings new challenges to PRNU-based image forensics. For example, the performance of the existing forensic methods may deteriorate due to different camera exposure parameter settings and the efficacy of the PRNU-based methods can be directly challenged by image editing tools from social network sites or anti-forensics attacks. The objective of this thesis is to investigate and design effective methods to mitigate some of these challenges on PRNU-based image forensics.

We found that the camera exposure parameter settings, especially the camera sensitivity, which is commonly known by the name of the ISO speed, can influence the PRNU-based image forgery detection. Hence, we first construct the Warwick Image Forensics Dataset, which contains images taken with diverse exposure parameter settings to facilitate further studies. To address the impact from ISO speed on PRNU-based image forgery detection, an ISO speed-specific correlation prediction process is proposed with a content-based ISO speed inference method to facilitate the process even if the ISO speed information is not available. We also propose a three-step framework to allow the PRNUbased source oriented clustering methods to perform successfully on Instagram images, despite some built-in image filters from Instagram may significantly distort PRNU. Additionally, for the binary classification of detecting whether an image's PRNU is attacked or not, we propose a generative adversarial network-based training strategy for a neural network-based classifier, which makes the classifier generalize better for images subject to unprecedented attacks.

The proposed methods are evaluated on public benchmarking datasets and our Warwick Image Forensics Dataset, which is released to the public as well. The experimental results validate the effectiveness of the methods proposed in this thesis.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item