Deepfake Videos: Detection Methods

Introduction

Context of the Problem

In the past few years, the issue of face-manipulated videos has received significant

attention. This has been so particularly due to the advent of deepfake technologies that can

manipulate videos and images using deep-learning tools. With the deepfake algorithm, faces

in the target video can be replaced with faces in the source video employing either generative

adversarial networks or autoencoders. With deepfake technology, it is possible to create face-

manipulated videos and photos, provided that the creator can access considerable amounts of

data. Although deepfake videos are not well understood by the public, they are constantly

creating fear among people. It is quite hard to trust what one reads or watches because it is

quite difficult to tell the difference between deep fakes and genuine videos. There is also

much concern, given that it can be used to destroy someone’s reputation and compromise

their integrity. Therefore, there is a concern on how to tell if a video is genuine or a deepfake.

Brief Introduction of the Problem

As with several developments in the history of man, deepfake videos have their

origins in pornography. Although the manipulation of images has been around for several

years, deepfakes came to life and public life in 2017 when a user on Reddit created deepfake

pornographic content videos featuring celebrities. In these videos, celebrity faces were

swapped into adult actors’ bodies. Although the group involved with this content was banned,

deepfakes around pornographic content continue to be developed and shared online. Since

these deepfakes featured in public, several organizations have established apps that rely on

this technology. Some of these apps may be much less advanced, producing less life-like

results. Deepfakes from such apps can be easily detected as such, meaning that there is no

problem detecting that the video is fake. However, some apps are quite advanced and are