Facial Action Coding System (FACS)

FACS defines strict anatomical constraints on which Action Units (AUs) can physically occur together. Generative models frequently violate these rules by producing impossible muscle combinations—such as activating the frontalis (AU 1) to raise the inner brow while simultaneously activating the corrugator supercilii (AU 4) in a pattern that contradicts human musculoskeletal structure.

• Key violation: Simultaneous AU 12 (lip corner puller) and AU 15 (lip corner depressor) at full intensity
• Detection approach: Build a co-occurrence matrix from ground-truth FACS-coded datasets and flag statistically impossible AU pairs
• Forensic value: These anatomical violations persist across different deepfake architectures and are difficult to patch without explicit physiological modeling

Natural facial expressions exhibit hemispheric asymmetry in AU intensity due to contralateral motor cortex innervation. Deepfake generators tend to produce unnaturally symmetric muscle activations, creating an uncanny uniformity that FACS-based intensity scoring can quantify.

• Measurement: Compare AU intensity scores (A–E scale) between left and right facial hemispheres
• Natural baseline: Genuine expressions typically show 5–15% intensity variance between sides
• Synthetic signature: Generated faces often show <2% variance, indicating algorithmic rather than biological motor control
• Advanced analysis: Track asymmetry dynamics across video frames to detect temporal smoothing artifacts

FACS coding captures the precise onset, apex, and offset timing of facial muscle movements. Micro-expressions—involuntary movements lasting 1/25 to 1/5 of a second—follow characteristic temporal envelopes that generative models struggle to reproduce.

• Attack phase: Natural AU activation follows a sigmoidal curve, not linear interpolation
• Decay phase: Muscle relaxation exhibits exponential decay with specific time constants per AU
• Synthetic failure mode: Generated faces often show constant-velocity transitions or abrupt frame-to-frame jumps
• Forensic metric: Measure the root-mean-square error between observed AU time courses and biologically validated temporal models

The Duchenne marker—contraction of the orbicularis oculi (AU 6) during genuine smiles—remains one of the most reliable forensic indicators. This muscle produces crow's feet wrinkles and raises the cheek, compressing the eye aperture in a way that requires modeling complex soft-tissue deformation.

• Genuine signature: AU 6 + AU 12 co-activation with lateral canthal lines (crow's feet)
• Synthetic failure: Deepfakes often produce AU 12 (lip corner pull) without corresponding AU 6 activation, creating a non-Duchenne or social smile
• Detection method: Track the ratio of eye aperture reduction to cheek elevation across frames
• Why it works: The orbicularis oculi's effect on periorbital skin texture requires high-frequency detail generation that current models smooth over

FACS provides the anatomical bridge between phonemes (speech sounds) and visemes (visual mouth shapes). Each phoneme requires specific AU combinations—for example, the /m/ phoneme requires AU 24 (lip presser) activation. Deepfake audio-visual synthesis frequently produces temporal misalignment or incorrect viseme selection.

• Mapping framework: Phoneme-to-AU lookup tables derived from linguistic phonetics
• Detection signal: Measure the temporal offset between audio envelope onset and corresponding AU activation
• Common error: Bilabial consonants (/p/, /b/, /m/) require complete lip closure (AU 24) that generators often render as partial occlusion
• Forensic pipeline: Extract MFCC features from audio, predict expected AU sequence, compare against observed facial AU time series

FACS codes lip separation with three hierarchically related AUs: AU 25 (lips part), AU 26 (jaw drop), and AU 27 (mouth stretch). These actions have strict anatomical dependencies—AU 27 cannot occur without AU 26, which cannot occur without AU 25. Generative models frequently violate this hierarchy.

• Anatomical rule: AU intensity must follow AU25 ≥ AU26 ≥ AU27
• Synthetic violation: Generated faces may show high-intensity mouth stretch (AU 27) without corresponding jaw drop (AU 26)
• Detection: Build a dependency graph of AU hierarchies and flag violations as synthetic indicators
• Persistence: This hierarchical violation appears across GAN-based and diffusion-based face generation methods

The automated detection of involuntary, fleeting facial muscle movements lasting 1/25th to 1/15th of a second. These expressions are difficult for synthetic face generation models to replicate with natural temporal dynamics.

• FACS provides the ground-truth coding framework for classifying these rapid movements
• Deepfake models often fail to produce the correct onset and offset timing of Action Units
• Genuine micro-expressions exhibit asymmetric activation patterns that GANs struggle to reproduce
• Used in high-stakes screening to detect concealed emotions that contradict displayed affect

A specific audio-visual forensic analysis that detects inconsistencies between spoken phonetic sounds (phonemes) and the observed mouth shapes (visemes) required to produce them.

• FACS Action Units AU 10, AU 12, AU 16, AU 20, AU 25, AU 26 directly map to specific viseme configurations
• Synthetic faces often produce impossible AU combinations for a given phoneme
• Temporal misalignment between audio onset and AU activation is a key deepfake indicator
• Analysis requires frame-accurate synchronization of audio and video streams

A deepfake detection metric measuring the temporal and spatial misalignment between visual lip movements and the corresponding audio speech track.

• FACS codes the precise intensity and duration of perioral Action Units (AU 10-AU 28)
• Deepfake generators often produce over-smoothed lip transitions lacking natural coarticulation
• Genuine speech exhibits anticipatory coarticulation where lip shapes prepare for upcoming phonemes
• Detection models compare predicted AU sequences from audio against observed visual AUs

A forensic technique that fits a three-dimensional face model to a two-dimensional image to detect inconsistencies in estimated shape, texture, and lighting parameters.

• FACS-based blendshapes are used to parameterize the expression component of the 3DMM
• Face-swapped deepfakes often produce inconsistent 3D shape estimates between facial regions
• The fitted model reveals impossible geometric configurations of Action Units
• Lighting and albedo parameters extracted from the 3DMM can be compared against FACS-driven expectations

A liveness detection method that extracts subtle skin color variations caused by blood flow from video to verify the presence of a living human cardiovascular system.

• FACS-defined facial regions (e.g., forehead, cheeks) serve as regions of interest for PPG signal extraction
• Synthetic faces lack the physiological blood volume pulse signal detectable in genuine skin
• Deepfake models often produce static skin texture without the micro-color variations of circulation
• Combining PPG with FACS AU detection creates a multi-modal liveness score

A video forensics method evaluating the coherence of motion, illumination, and texture across consecutive frames to identify frame-by-frame manipulation or insertion.

• FACS provides the biomechanical constraints for natural facial movement trajectories
• Genuine AU activations follow sigmoid-like onset and offset curves with specific acceleration profiles
• Deepfake-generated expressions often exhibit temporal jitter and non-physical AU co-activation patterns
• Frame-to-frame AU intensity deltas exceeding physiological limits indicate synthetic manipulation