[Paper Review] DeepFakesON-Phys: DeepFakes Detection based on Heart Rate Estimation

Javier Hernandez‐Ortega, Rubén Tolosana+2 more|arXiv (Cornell University)|Oct 1, 2020

Non-Invasive Vital Sign MonitoringEngineering38 references57 citations

TL;DR

DeepFakesON-Phys detects DeepFakes by estimating heart-rate related physiological signals from face videos using a Convolutional Attention Network, achieving state-of-the-art AUC on Celeb-DF v2 and DFDC Preview.

ABSTRACT

This work introduces a novel DeepFake detection framework based on physiological measurement. In particular, we consider information related to the heart rate using remote photoplethysmography (rPPG). rPPG methods analyze video sequences looking for subtle color changes in the human skin, revealing the presence of human blood under the tissues. In this work we investigate to what extent rPPG is useful for the detection of DeepFake videos. The proposed fake detector named DeepFakesON-Phys uses a Convolutional Attention Network (CAN), which extracts spatial and temporal information from video frames, analyzing and combining both sources to better detect fake videos. This detection approach has been experimentally evaluated using the latest public databases in the field: Celeb-DF and DFDC. The results achieved, above 98% AUC (Area Under the Curve) on both databases, outperform the state of the art and prove the success of fake detectors based on physiological measurement to detect the latest DeepFake videos.

Motivation & Objective

• Motivate the use of physiological measurements, specifically heart rate, to detect DeepFakes.
• Develop a DeepFake detector that transfers knowledge from heart-rate estimation to detection.
• Leverage rPPG-based features within a CAN architecture to improve robustness to visual artifacts.

Proposed method

• Base the detector on DeepPhys, adapting it for classification rather than heart rate estimation.
• Initialize with weights pretrained for heart rate estimation and fine-tune only the last layers.
• Two-branch Convolutional Attention Network (Motion and Appearance Models) share information through attention masks.
• Replace the original regression output with a sigmoid classification for real/fake detection at frame level.
• Preprocess frames with light normalization and use a frame-difference input for the Motion Model.

Experimental results

Research questions

• RQ1 Can rPPG-based features learned for heart rate estimation be repurposed to detect DeepFakes?
• RQ2 Does a CAN architecture effectively fuse spatial and temporal cues for physiological-based DeepFake detection?
• RQ3 How well does the proposed method generalize to unseen DeepFake databases (Celeb-DF v2 and DFDC Preview)?

Key findings

• DeepFakesON-Phys achieves AUC 99.9% on Celeb-DF v2 and 98.2% on DFDC Preview.
• Accuracy is 98.7% on Celeb-DF v2 and 94.4% on DFDC Preview.
• The detector is trained with frozen convolutional layers initialized from a heart-rate estimation model and only the final layers are trained.
• The approach outperforms state-of-the-art detectors based on Face Warping, LCN/mesoscopic features, and pure deep-learning features.
• The method demonstrates that current DeepFakes do not exploit heart-rate related physiological information in synthesis.