| dc.description.abstract | Face Recognition Systems (FRS) have a crucial role in various critical systems in modern society, making them targets for a range of attacks. This thesis explores the impact of a form of facial image manipulation called the Thatcher illusion, which inverts the eyes and mouth of faces on state-of-the-art FRS models and the performance of human observers’ ability to detect these manipulated images.
A diverse dataset of Thatcherized images was created to assess the performance of the FRS models. A platform for conducting eye-tracking platform and methods to analyze the collected data was developed. The cognitive processes of human observers were examined using this dataset. The performance of observers with and without providing a reference image while detecting the manipulated images was analyzed along with the eye fixation over different regions of the facial images by human observers during these experiments.
The findings suggest that FRS models are vulnerable to these types of low-effort manipulations. The overall manipulation detection accuracy of observers was observed to vary significantly with the presence of reference images, however, the kinds of classification errors made by observers differed based on these factors. Additionally, the gender of human observers had a significant relation to their performance. The order in which manipulation detection is done with and without the reference image was also found to have a significant relation to the performance. While detecting manipulated images, observers were found to have a higher proportion of fixation on the nose and mouth regions of the facial images. The findings highlight the need to consider a diverse set of factors that affect human cognition and competence in image manipulation detection.
The thesis contributes towards understanding the vulnerability of FRS against low-effort manipulation techniques like Thatcher illusion and presents valuable insights about human competence for detecting such manipulations. | |