Deepfake images detection for forensic purposes

Issues arise from the production and circulating of high-fidelity synthetic images of people, also known as deepfake, have been in the headline of news. The potential of using these images for disinformation has lead politicians to consider criminalising distribution of non-consensual deepfake images (Hern 2022). As these images are pixel-wise generated using deep learning methods, traditional image processing based detection techniques such as identifying non-continuous edges, uneven noise distributions and illumination variations within images are ineffectual (Wong 2019). Social media platforms have a particular interest in detecting deepfake content as they can be sued or heavily fined if they do not remove these images in a timely manner. As a result, social media platforms have been researched and sponsored research on deepfake detection (Deepfake Detection Challenge 2019) and some successes have been achieved. However, as the techniques for generating deepfake keep evolving accurate detection of deepfake imagery is still a challenge and a fertile research area. Furthermore, current research mainly focuses on the technical aspects of deepfake detection, while the forensic aspects have not been fully considered. If a claimed deepfake image is presented in court as evidence, the integrity of the procedures and processes for the detection need to be validated and explained. The aim of this study is thus to develop a deepfake detection framework that is suitable for forensic purposes.

Objectives

The main objectives of this project are:

To critique recent literature on the creation and detection of deepfake imagery and to frame the existing research gaps
To identify or develop a rich dataset for evaluating the effectiveness of deepfake detection methods
To design, develop, test and refine the deep-fake detection method, which achieves high detection accuracy and provide explanation of the detection outcomes