Year: 2023
Type: dissertation
Abstract: The term “deepfakes” refers to media content that has been manipulated using deep learning. This thesis project seeks to answer the question of how well humans are able to detect deepfakes. In particular, the project compares people’s ability to detect deepfakes between two different deepfake categories; face swap and lip sync. In order to achieve this, a perceptual user test was performed, in which 30 participants were given a number of lip sync, face swap and unaltered videos and were asked to classify which of them were unaltered and which of them were manipulated using deepfake technology. These results serve to fill in the gap in knowledge regarding perceptual user tests on deepfakes, for which only a small amount of research has been made. The results also serve to shed light on which types of deepfakes pose the biggest threat regarding the problem of malicious impersonation. The main conclusion from this study was that lip sync is likely harder for humans to detect than face swap. The percentage of correct classifications of lip sync videos was 52.7%, and the percentage of correct classifications of face swap videos was 91.3%.
Author: Sundström, Isak
Institution: KTH Royal Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6980938987
