Model and Physiological based Deep Fake Detection

"The rapid proliferation of deepfake videos poses a growing threat to the integrity of digital information and public discourse. Traditional detection methods, often based on convolutional neural networks or artifact-specific analyses, struggle to keep up with the fast pace of generative model improvements. This work explores the use of physiological and behavioral signals extracted from facial videos (such as heart rate (rPPG), blink dynamics, and gaze behavior), for deepfake detection. We developed an open-source system called Micro-signals and Involuntary Cues for Unmasking Synthetics (MICUS), which implements modules for signal extraction and classification based on handcrafted features. Using a balanced subset of 509 videos (232 real and 277 fake) sourced from the Celeb-DF and FaceForensics++ datasets, our best model (Gradient Boosting with feature selection and PCA) achieved an accuracy of 72.55%. The results highlight the potential of interpretable, biologically grounded features as complementary indicators of facial manipulation".