Validation of PRNU source-camera identification error rates using methods and statistics on dark and bright images
Date
2024-08
Authors
Martin, Abby Joanne
Major Professor
Advisor
Newman, Jennifer
Tian, Jin
Ommen, Danica
Catanzaro, Michael
Maxion, Roy
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Source camera identification is a field of digital image forensics which aims to match a questioned image with the camera that took it. The photo response non-uniformity (PRNU) camera fingerprint is the basis for a source-camera identification algorithm used in court. PRNU is a persistent noise which is unique to each camera sensor and is included in every image taken by that camera. Our work aims to determine how overall image brightness impacts the algorithm's error rates to determine whether these differences are due to chance, a previously uninvestigated area. We use two datasets: our own curated collection of images, the StegoAppDB dataset, with ISO and exposure time settings intentionally offset from the auto-exposure settings to produce nominal, dark, and bright images, and a set of images downloaded from Flickr. For the PRNU source-camera identification algorithm applied to the StegoAppDB data, false negative rates increase for under- and over-exposed questioned images by about 15% compared with auto-exposed questioned images, and the false positive rate is approximately one in 200 for under-exposed images. To apply the camera identification algorithm to questioned images with unknown exposure type, we propose a classification method to label any image as nominal, dark, or bright based on its complementary cumulative mass function (CCMF). Visual properties of an image can be subjective; therefore, we validate the exposure type of the ground-truth data for our classification method using collection settings, exposure values, and through an interrater agreement study between three human judges and the computer labeling, and find all three validation methods have close results. We show on both data sets that the false negative and false positive rates for dark and bright labeled questioned images can differ significantly from the error rates for nominal images. We use hypothesis tests to validate these differences in error rates between exposure classes for the PRNU source-camera identification algorithm, and confirm these differences are statistically significant and not due to chance. In particular, we find that dark or bright questioned images can result in different false negative rates than nominal questioned images, and that the false positive rates differ for the dark and nominal questioned images. We introduce the use of hypothesis testing to determine how the error rates of an established forensic tool can change for different image features, specifically image brightness levels.
Series Number
Journal Issue
Is Version Of
Versions
Series
Academic or Administrative Unit
Type
dissertation