The effect of a little amount of training based on current artificial deepfake detection methods on the ability of the general public to differentiate between deepfakes and non-deepfakes: a pilot study
Keywords
Loading...
Authors
Issue Date
2021-02-09
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Creating deepfakes has in recent years changed from being technically hard to accomplish to being easily doable by the general public with the use of open-source software, computer programs, and even mobile phone applications. This development has made it possible for even more people to do harm with this kind of software when using it to deceive other people. This thesis investigates whether it is possible to increase the ability of the general public to differentiate between deepfakes and non-deepfakes by providing a small training of only a few minutes based on certain deepfake characteristics. These deepfake characteristics are based on what current artificial deepfake detection methods train upon. In this way, humans would be educated by the debunking of artificial deepfake detection methods, to increase their ability to differentiate between deepfakes and non-deepfakes. This is tested in an experiment that is based on the provided literary foundation. The results showed a non-significant difference between the ability of the control group and the ability of the experimental group to differentiate between deepfakes and non-deepfakes when all data of all participants was taken into account (p = 0.059). When only taking into account the participants who were only a bit familiar or not familiar at all with deepfakes, the results did show a significant difference between the ability of the control group and the experimental group to differentiate between deepfakes and non-deepfakes (p = 0.042). The sample size used in this thesis is too small to generalize its findings to the total population. Performing the experiment with a larger set of participants might result in more conclusive results and conclusions concerning the total population. When the training shows to help significantly when performed with a larger set of participants, this training of only a few minutes could be distributed to the general public. This could, for example, be done in the form of a national public information advertising campaign, to reduce the amount of harm caused as a result of deception by deepfakes. A national public information advertising campaign would be a great way to combat the dangerous implications of deepfakes in addition to just using artificial deepfake detection methods. There should be especially focused upon people who are only a bit familiar or not at all familiar with deepfakes. The reason for this is that in the experiment of this thesis, the ability of this subgroup to differentiate between deepfakes and non-deepfakes was significantly better when receiving the training of only a few minutes. These results might also particularly show when the experiment is performed with a larger set of participants.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen