On Using rPPG Signals for DeepFake Detection: A Cautionary Note

Alessandro D'Amelio, Raffaella Lanzarotti, Sabrina Patania, Giuliano Grossi, Vittorio Cuculo, Andrea Valota, Giuseppe Boccignone

Abstract: An experimental analysis is proposed concerning the use of physiological signals, specifically remote Photoplethysmography (rPPG), as a potential means for detecting Deepfakes (DF). The study investigates the effects of different variables, such as video compression and face swap quality, on rPPG information extracted from both original and forged videos. The experiments aim to understand the impact of face forgery procedures on remotely-estimated cardiac information, how this effect interacts with other variables, and how rPPG-based DF detection accuracy is affected by these quantities. Preliminary results suggest that cardiac information in some cases (e.g. uncompressed videos) may have a limited role in discriminating real videos from forged ones, but the effects of other physiological signals cannot be discounted. Surprisingly, heart rate related frequencies appear to deliver a significant contribution to the DF detection task in compressed videos.