Robust Deepfake Detection using a 1D Convolution-Transformer Hybrid Model based on Remote Photoplethysmography Signals
Oral presentation
Research approach
Isolated skin regions with BiSeNet and extracted rPPG with PhysNet. A 1D CNN captures local temporal features and a Transformer models global dependencies to classify real and synthetic videos.
View full-size figure
Results
Experiments used FaceForensics++ and CelebV-HQ with a 6:2:2 train/validation/test split. Three convolutional blocks reached 88.58% accuracy, compared with 84.95% for the Transformer alone.
All publications