Towards Robust Deepfake Detection Based on Heart Rate Analysis

• Soo-Hyun Lee,
• Gyung-eun Yun,
• Seong Hee Park,
• Min Young Lim,
• Youn Kyu Lee,

Vol. 19, No. 1, January 31, 2025
10.3837/tiis.2025.01.009, Download Paper (Free):

• Deepfake
• Deepfake detection
• remote photoplethysmography
• rPPG
• Face Recognition

Abstract

Remote photoplethysmography (rPPG)-based deepfake detection analyzes a target’s heart rate (HR) patterns in a video. However, existing studies on rPPG-based deepfake detection have primarily focused on detection algorithms that utilize extracted HR values, rather than on mechanisms for acquiring and analyzing informative HR patterns. To establish robust rPPG-based deepfake detection, it is required to clarify the impact of key factors that directly affect HR extraction and processing. We defined the key factors that directly affect the performance of rPPG-based deepfake detection, including the facial region, extraction interval, rPPG extraction method, and feature engineering method. Based on these defined key factors, we assessed the impact of each variable for the key factors on rPPG-based deepfake detection performance. Furthermore, we identified the optimal combination of variables, which maximizes the performance of rPPG-based deepfake detection. Through statistical validation of 164 real-world videos, we conducted experimental evaluations. The results revealed variables for each key factor that yielded significant differences in HR values between real and fake videos. Moreover, we identified the optimal combination of the variables, enabling a pronounced distinction between real and fake videos. This paper successfully confirms the impactful role of individual variables and their optimal combinations for each key factor in rPPG-based deepfake detection.

Statistics

Cite this article