Using Eye-tracking to Study the Authenticity of Images Produced by Generative Adversarial Networks
Nicholas Caporusso, Kelei Zhang, Gordon S. Carlson · 2020 International Conference on Electrical, Communication, and Computer Engineering (ICECCE) · 2020
Nowadays, Machine Learning algorithms, such as Generative Adversarial Networks (GANs), enable generating content, and especially images, featuring people, objects, or landscapes, with unprecedented levels of accuracy and fidelity. As a result, it is becoming challenging for a viewer to distinguish a picture of a fake profile from one that has a real human in it. In this paper, we present the results of an experimental study in which we investigated the perception of images produced by GANs. Specifically, we focused on the individuals' ability to discriminate between fake and real profiles. Furthermore, we utilized eye-tracking technology to identify the presence of patterns in subjects' gaze, which, in turn, can be useful to optimize the output of GANs and, simultaneously, provide insight on the underlying cognitive dynamics.