Detection of Photoplethysmography Manipulation in Video Forgery

Recent studies have shown that physiological signals related to blood pressure and heart rate can be estimated in a contactless modality from facial videos using remote photoplethysmography (rPPG). This has paved the way to the development of techniques that can acquire and manipulate the rPPG signals recoverable from facial videos without affecting their visual appearance. The goal of this paper is to analyze the detectability of this new kind of forgery, here referred to as rPPG deepfake videos. Specifically, we propose a two-stream method based on the analysis of rPPG deepfake videos at both spatial and temporal levels. The experimental results obtained from tests performed on samples taken from two distinct databases demonstrate that our method performs better than popular deep learning methods in the rPPG deepfake detection task.