Micro-Expression Analysis

The foundational process of identifying individual facial muscle activations as defined by the Facial Action Coding System (FACS). Automated classifiers map pixel-level deformations to specific AUs, such as AU4 (Brow Lowerer) or AU12 (Lip Corner Puller).

• Detects 40+ anatomically distinct muscle groups
• Temporal resolution: < 40ms window for true micro-expressions
• Synthetic faces often fail to co-activate antagonist muscles naturally

Measures the onset, apex, and offset phases of a facial movement. Genuine micro-expressions follow a smooth, non-linear velocity profile governed by muscle physiology.

• Natural onset-to-apex duration: 100-200ms
• Synthetic faces exhibit linear or jittery motion interpolation
• Optical flow consistency across frames is a key discriminative feature

Involuntary emotional expressions often manifest with subtle bilateral asymmetry. Deepfake generators tend to produce unnaturally perfectly symmetrical muscle activations.

• Measures left-right intensity differentials per AU
• Spontaneous smiles show 5-15% asymmetry; posed/synthetic smiles approach 0%
• Specular highlight mismatch on eyes can corroborate asymmetry findings

FACS defines anatomically impossible or highly improbable action unit combinations. A synthetic face might simultaneously activate AU5 (Upper Lid Raiser) and AU43 (Eye Closure), a physically contradictory state.

• Validates AU combinations against kinesiological constraints
• Flags non-Duchenne markers (e.g., smile without orbicularis oculi engagement)
• Used in conjunction with Phoneme-Viseme Mismatch analysis for video

Micro-expressions are temporally correlated with subtle skin color variations caused by blood flow. Remote PPG extracts heart-rate signals from facial video; synthetic faces lack a genuine cardiovascular pulse signal.

• Extracts chrominance changes invisible to the human eye
• Validates that expression dynamics align with a living circulatory system
• Counters 3D mask and hyper-realistic avatar presentation attacks

Fitting a 3D Morphable Model (3DMM) to a 2D face image recovers shape, texture, and lighting parameters. Micro-expression analysis constrains the 3DMM's blend shapes to only anatomically valid deformations.

• Detects impossible 3D shape configurations in deepfake geometry
• Validates that expression-driven surface deformations respect muscle attachment points
• Integrates with Lighting Inconsistency Analysis for holistic forgery detection

The foundational anatomical taxonomy for micro-expression analysis, FACS defines 44 distinct Action Units (AUs) corresponding to individual facial muscle contractions. Each AU is coded for onset, apex, offset, intensity, and asymmetry. Deepfake detection systems use FACS as a ground-truth framework: generative models frequently produce impossible AU combinations—such as activating the zygomaticus major (AU12) without the orbicularis oculi (AU6) in a genuine Duchenne smile—or exhibit non-biological co-activation patterns where antagonistic muscle groups fire simultaneously. Automated FACS coding via computer vision enables frame-by-frame comparison against anatomical plausibility constraints.

Micro-expressions exhibit a characteristic temporal signature that generative models struggle to replicate. Authentic expressions follow a three-phase trajectory: onset (rapid muscle contraction over 20-50ms), apex (peak intensity held for <100ms), and offset (asymmetric relaxation). Synthetic faces often display linear interpolation between expression states rather than the non-linear, ballistic muscle dynamics of real faces. Key forensic metrics include:

• Rise time: Duration from neutral to peak intensity
• Hold duration: Time spent at maximum deformation
• Fall time: Asymmetry ratio between left and right hemiface relaxation
• Jitter: Frame-to-frame micro-fluctuations absent in smooth synthetic generation

A complementary liveness verification technique that extracts subtle skin color variations caused by cardiac blood flow from standard video. The human cardiovascular system produces imperceptible chromatic changes—typically 0.5-1% pixel intensity variation at frequencies of 0.7-4 Hz—that are absent in synthetic face generation. Deepfake models do not model hemodynamic physiology, resulting in flat PPG signals when extracted from synthetic facial regions. Forensic systems combine PPG analysis with micro-expression detection to create a multi-modal liveness score: a genuine face must simultaneously exhibit anatomically correct muscle movements and physiological blood flow signals.

A rule-based forensic layer that encodes the biomechanical interdependencies of facial muscles. Certain AU pairs are anatomically impossible or statistically improbable in genuine expressions due to shared musculature or neural innervation pathways. For example:

• AU4 (brow lowerer) and AU1 (inner brow raiser) rarely co-activate at high intensity due to antagonistic muscle groups
• AU10 (upper lip raiser) and AU15 (lip corner depressor) require independent levator and depressor control that synthetic models often conflate
• Unilateral AU activation without contralateral mirroring violates facial nerve innervation patterns Detection systems maintain a co-occurrence matrix derived from large corpora of genuine expressions, flagging statistically aberrant AU combinations as potential synthetic artifacts.

A specialized audio-visual analysis that detects inconsistencies between spoken phonetic sounds and the corresponding visual mouth shapes required to produce them. Each phoneme maps to a specific viseme—the visual equivalent of a phoneme—defined by lip rounding, jaw aperture, and tongue visibility. Deepfake systems often generate plausible but incorrect viseme sequences because they model speech-driven animation as a statistical mapping rather than a biomechanical articulation process. Forensic systems extract phoneme alignments from the audio track using forced alignment and compare the predicted viseme sequence against the observed facial landmarks frame-by-frame, flagging temporal misalignments exceeding 40ms.

A motion estimation technique that computes the dense displacement vector field between consecutive video frames to characterize facial movement patterns. Authentic micro-expressions produce smooth, biologically constrained flow fields with gradual spatial gradients. Synthetic faces often exhibit:

• Quantized motion vectors from discrete latent space interpolation
• Uniform flow magnitude across facial regions that should move at different rates
• Temporal flicker where flow direction reverses unnaturally between frames
• Boundary discontinuities at the face-background interface where motion vectors terminate abruptly Optical flow analysis serves as a low-level motion forensics layer that operates independently of expression classification, detecting synthetic artifacts even when the high-level expression appears semantically correct.