An Overview of Deepfake Video Detection Using Remote Photoplethysmography

Deepfake technology, which can create remarkably realistic videos through deep learning techniques, has many applications that serve humanity, including cinema and television productions, education, social media applications, art, fashion, and virtual assistants. However, this technology also brings with it potential abuse scenarios through manipulative content. Particularly in cases like fake news, identity fraud, blackmail, and slander, it can hasten the dissemination of false information and violate people’s privacy. Moreover, it may cause major legal and societal issues in crucial fields like politics and public security. In this context, developing solutions for detecting deepfake videos has become imperative. While some methods developed in the literature to detect deepfake video are based on spatial and frequency-based analysis of digital traces and residues, which emerge intrinsically during the fake content production stage, some are based on examining physiological signs. Remote photoplethysmography (rPPG)-based physiological approaches, which analyze imperceptible color changes on the skin surfaces of individuals, have gained significant attention due to their high performance. This study examined rPPG-based techniques and their effectiveness in detecting fake videos made with Generative Adversarial Network (GAN) and Autoencoder (AE), two of the most popular deep learning algorithms used to produce deepfake content, and discussed the technical challenges encountered.