"Deepfakes" refer to fake media, like pictures and videos, that are developed using deep neural networks. Unlike fake media created using Photoshop, these forgeries are almost indistinguishable from the real thing.

In this project, we explore the world of Deepfakes using a pre-trained model engineered to detect them, known as MesoNet.

The MesoNet research paper was published by Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen in September 2018. The paper includes two models each trained on two different datasets: we explore the Meso4 model trained on the Deepfake dataset, which was created and released by MesoNet researchers.

• Paper: MesoNet: a Compact Facial Video Forgery Detection Network
• GitHub repo: DariusAf/MesoNet

Deepfake research rapidly evolves, and we suggest keeping up with the latest research using tools like Google Scholar

(This exploration was developed in partnership with Mikhail Lenko.)