Deepfake Video Detection Using the Frequency Characteristic of Remote Photoplethysmography

Abstract

Photoplethysmography is a technique for measuring the blood flow per unit time of an artery. Remote photoplethysmography is a method for obtaining photoplethysmography signals in a non-contact manner through a sensor such as a camera and has been recently applied to various fields. In this study, we propose a method for detecting Deepfake modulated color video based on remote photoplethysmography concept. As experimental data, 50 real videos and their 50 Deepfake videos using Face Swapping Generative Adversarial Networks were used. The photoplethysmography signals of face and neck regions were extracted, respectively, and the signals were preprocessed by detrending and performing Butterworth bandpass filtering. The 80 power values ​​in the frequency domain were defined as feature vectors. As a result of analyzing the L2 Norm between the two vectors extracted from the face region and the neck region, the L2 Norms of the real video and the fake video were 0.0000307 and 0.0001332, respectively, confirming that the distributions were clearly separated. It was confirmed that there is a significant difference between the real and the fake videos. Also, as a result of calculating the degree of separation of distributions with d-prime, 2.32 was derived.