Matthew Yates (2018) – thesis

Full e-thesis links are provided below. If needed, theses can also be found in the University of Nottingham e-Theses Repository.

Accurate detection methods for GAN-generated earth observation images using expert visual perception

Dr Matthew Yates (2018 cohort)

Abstract:

Image generation techniques, such as generative adversarial networks (GANs), have become sufficiently sophisticated to cause growing security concerns regarding image authenticity. Although generation and detection methods are often applied to a range of images such as objects and faces, more domain specific image types such as Earth Observation (EO) have received relatively little attention, leaving the field vulnerable to potential malicious misuse of this technology. This thesis investigates the current state of EO specific GAN generation and detection methods using an interdisciplinary approach. This work argues that further detection methods should incorporate both human and computational detection to improve current techniques. Evidence to support this conclusion is given by the following contributions:

1. A literature review of the current state of image generation and detection with respect to EO imagery.
2. A new benchmark evaluation of current GAN models in the task of the unconditional generation of synthetic EO imagery.
3. A Comparison between detection methods in both human and computer detection systems towards synthetic EO imagery that quantifies the key behavioural differences and effectiveness for each approach. The findings from two image detection studies show that these systems prioritize different image features for making accurate detections.
4. An eye-tracking image detection study between expert and novice users. The results find that experts exhibit more efficient and effective visual search strategies for detection.
5. The development of a novel framework to improve current techniques by guiding a CNN detection model using eye gaze data from self-reported high experience individuals. The results found that this approach increased detection performance over control models.

School of Computer Science

Supervisors: Michael Pound, Robert Houghton

Keywords: Machine learning, Generative adversarial networks, Deep Learning, Synthetic image detection, Eye tracking, Earth observation imagery, Image generation