Biological Signals¶
Biological signals are implicit measurements of physiological processes visible in human facial images and videos—including blood flow patterns (photoplethysmography), eye blinks, micro-expressions, and heart rate. These signals are difficult for generative models to synthesize realistically, making them valuable for forensic analysis and authenticity verification.
Key physiological signals¶
- Photoplethysmography (PPG): Measurement of blood flow variation in facial capillaries via reflectance changes in video frames. Can be measured via green-channel PPG (G-PPG) or chrominance-based PPG (C-PPG).
- Eye blink patterns: Natural blinking follows predictable statistical patterns that synthetic faces often fail to replicate accurately.
- Micro-expressions: Brief, involuntary facial expressions lasting 0.5–4 seconds that GANs struggle to generate naturally.
- Heart rate variability: Temporal patterns in heart rate extracted from subtle facial color changes.
Why biological signals matter for detection¶
Generative models (GANs, etc.) learn to reconstruct visual appearance but do not inherently understand or preserve the coherence of underlying physiological processes. Key properties:
- Spatial coherence: Biological signals extracted from different facial regions (cheeks, forehead) are correlated in authentic videos but often inconsistent in synthetic videos.
- Temporal consistency: Authentic biological signals exhibit stable statistical properties across video frames; synthetic videos often exhibit discontinuities.
- Cross-modal consistency: Audio-visual synchronization (lip movements matching speech) and heart rate coherence with visible chest motion are hard to fake.
These properties are orthogonal to visual artifacts (compression, lighting, frequency-domain anomalies), providing complementary detection signals.
Key papers¶
- FakeCatcher: Detection of Synthetic Portrait Videos using Biological Signals — Detection of deepfakes using photoplethysmography; demonstrates 91%+ accuracy by analyzing spatial coherence and temporal consistency of PPG signals
- In Ictu Oculi: Exposing AI Generated Fake Face Videos by Detecting Eye Blinking — Eye blink detection for deepfake identification; shows natural blinking patterns absent in synthetic faces
Detection methods¶
- PPG-based: Extract green-channel or chrominance PPG from facial regions; analyze spatial correlations and frequency-domain properties
- Blink-based: Detect frame-by-frame eye closure sequences; compare statistics to natural blinking models
- Heart rate based: Extract subtle color variations to estimate heart rate; synthetic videos often fail to maintain physiologically plausible rates
- Micro-expression: Detect fine-grained facial movements inconsistent with coarse-level manipulations