Temporal Image Forensics for Picture Dating based on Machine Learning

Ahmed, Farah (2022) Temporal Image Forensics for Picture Dating based on Machine Learning. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) ahmed.farah_phd_15036105.pdf - Submitted Version Download (2MB) | Preview

Abstract

Temporal image forensics involves the investigation of multi-media digital forensic material related to crime with the goal of obtaining accurate evidence concerning activity and timing to be presented in a court of law. Because of the ever-increasing complexity of crime in the digital age, forensic investigations are increasingly dependent on timing information. The simplest way to extract such forensic information would be the use of the EXIF header of picture files as it contains most of the information. However, these header data can be easily removed or manipulated and hence cannot be evidential, and so estimating the acquisition time of digital photographs has become more challenging.
This PhD research proposes to use image contents instead of file headers to solve this problem. In this thesis, a number of contributions are presented in the area of temporal image forensics to predict picture dating. Firstly, the present research introduces the unique Northumbria Temporal Image Forensics (NTIF) database of pictures for the purpose of temporal image forensic purposes. As digital sensors age, the changes in Photo Response Non-Uniformity (PRNU) over time have been highlighted using the NTIF database, and it is concluded that PRNU cannot be useful feature for picture dating application. Apart from the PRNU, defective pixels also constitute another sensor imperfection of forensic relevance. Secondly, this thesis highlights the fact that the filter-based PRNU technique is useful for source camera identification application as compared to deep convolutional neural networks when limited amounts of images under investigation are available to the forensic analyst. The results concluded that due to sensor pattern noise feature which is location-sensitive, the performance of CNN-based approach declines because sensor pattern noise image blocks are fed at different locations into CNN for the same category. Thirdly, the deep learning technique is applied for picture dating, which has shown promising results with performance levels up to 80% to 88% depending on the digital camera used. The key findings indicate that a deep learning approach can successfully learn the temporal changes in image contents, rather than the sensor pattern noise.
Finally, this thesis proposes a technique to estimate the acquisition time slots of digital pictures using a set of candidate defective pixel locations in non-overlapping image blocks. The temporal behaviour of camera sensor defects in digital pictures are analyzed using a machine learning technique in which potential candidate defective pixels are determined according to the related pixel neighbourhood and two proposed features called local variation features. The idea of virtual timescales using halves of real time slots and a combination of prediction scores for image blocks has been proposed to enhance performance. When assessed using the NTIF image dataset, the proposed system has been shown to achieve very promising results with an estimated accuracy of the acquisition times of digital pictures between 88% and 93%, exhibiting clear superiority over relevant state-of-the-art systems.

Actions (login required)

View Item

Downloads