Deepfake Detection Provenance

Detects synthetic media by analyzing biological signals that generative models fail to replicate authentically.

• Heart Rate Variability: Remote photoplethysmography (rPPG) extracts subtle skin color changes to detect pulse patterns absent in deepfakes.
• Eye Movement: Tracks saccadic patterns and blink rates; synthetic faces often exhibit unnatural gaze consistency or missing micro-saccades.
• Breathing Patterns: Analyzes chest movement periodicity; generative models struggle with physiologically accurate respiratory rhythms.

Real-world example: DARPA's Media Forensics program uses these signals to achieve >95% detection accuracy on high-quality deepfakes.

Identifies model-specific artifacts left by the generation process, enabling both detection and attribution to a specific AI architecture.

• GAN Fingerprints: Analyzes spectral discrepancies in frequency domain; each GAN architecture leaves a unique "device fingerprint" in generated images.
• Diffusion Model Traces: Detects characteristic noise residual patterns from the iterative denoising process.
• Checkerboard Artifacts: Identifies pixel-level grid patterns from transposed convolution layers in decoder networks.

These fingerprints function like ballistic markings on a bullet, linking synthetic media to its generative model of origin.

Examines media in the Fourier domain where generative models often fail to reproduce natural frequency distributions.

• Power Spectrum Anomalies: Real images follow a 1/f power law; GAN-generated images show characteristic peaks at specific frequencies.
• DCT Coefficient Analysis: JPEG compression domain reveals statistical inconsistencies in high-frequency components.
• Wavelet Decomposition: Multi-scale analysis exposes blending boundaries between synthesized and authentic regions.

This technique is particularly effective because upsampling operations in generators leave distinctive frequency-domain signatures invisible to human eyes.

Evaluates frame-to-frame coherence in video to detect the temporal instability characteristic of frame-by-frame generation.

• Optical Flow Inconsistency: Measures motion vectors between frames; deepfake videos often show unnatural warping artifacts.
• Flicker Detection: Identifies high-frequency luminance changes at face boundaries where blending fails.
• Identity Vector Drift: Tracks facial embedding consistency; synthetic faces may subtly shift identity across frames.

Temporal analysis catches what spatial-only detectors miss: perfect individual frames that fail to cohere into a natural video sequence.

Trains forensic classifiers to identify not just that content is synthetic, but which specific model architecture or training instance generated it.

• Closed-Set Attribution: Maps artifacts to known model families (StyleGAN, Stable Diffusion, DALL-E).
• Open-Set Verification: Determines if content originated from a previously unseen model.
• Training Instance Tracing: Identifies whether two synthetic media samples came from the exact same trained model checkpoint.

This enables forensic accountability, allowing investigators to trace deepfakes back to specific generation pipelines and potentially their operators.

Combines multiple forensic signals across modalities to achieve robust detection resistant to adversarial countermeasures.

• Audio-Visual Synchronization: Detects mismatches between lip movements and speech; deepfakes often show subtle audio-visual asynchrony.
• Cross-Modal Embedding: Projects visual and audio features into a joint space where authentic pairs cluster together.
• Ensemble Decision Logic: Fuses physiological, frequency, and temporal detectors with weighted voting to resist single-signal evasion.

Fusion systems achieve state-of-the-art robustness because attackers must simultaneously defeat multiple independent detection vectors.

Detects synthetic media by analyzing biological signals that generative models fail to replicate accurately:

• Heart Rate Variability: Extracts subtle skin color changes caused by blood flow (photoplethysmography) to detect absent or irregular pulse signals in deepfake faces
• Eye Movement Patterns: Analyzes saccadic motion, blink rates, and corneal reflections—synthetic faces often exhibit unnatural gaze consistency or missing physiological micro-movements
• Respiratory Motion: Detects chest and shoulder movements inconsistent with natural breathing cycles

These signals are computationally invisible to most generators, making them robust forensic markers.

Identifies model-specific artifacts left in synthetic media that serve as a forensic signature of the generating architecture:

• GAN Fingerprints: Analyzes spectral discrepancies in frequency domains unique to generative adversarial network upsampling operations
• Diffusion Model Traces: Detects characteristic noise residual patterns from the iterative denoising process
• Transformer Artifacts: Identifies statistical anomalies in token distribution and attention pattern signatures in generated text

Each model family leaves a distinct fingerprint enabling attribution to specific architectures or even individual training runs.

Examines media in the frequency domain rather than spatial domain to reveal hidden generation artifacts:

• Discrete Fourier Transform Analysis: Reveals periodic anomalies in pixel correlations invisible to human eyes
• Wavelet Decomposition: Isolates inconsistencies across multiple resolution scales typical of neural upscaling
• Power Spectrum Anomalies: Detects unnatural energy distributions in high-frequency bands where generators struggle

Frequency analysis remains effective even against adversarial perturbations designed to fool spatial-domain detectors.

Evaluates frame-to-frame coherence to expose synthetic video that lacks natural temporal dynamics:

• Optical Flow Inconsistency: Detects physically impossible motion vectors between consecutive frames
• Flicker Artifacts: Identifies inconsistent lighting, shadows, or texture stability across frames
• Lip-Sync Mismatch: Measures phoneme-viseme alignment errors where mouth movements don't match audio

Generative models often process frames independently, creating temporal discontinuities absent in real footage.

Traces synthetic media back to its source model or training dataset through forensic markers:

• Architecture Fingerprinting: Classifies the generating model family (StyleGAN, Stable Diffusion, DALL-E) by its unique artifact profile
• Training Data Membership Inference: Determines if specific images were in the model's training set through shadow model techniques
• Version Identification: Distinguishes between model checkpoints by subtle behavioral differences in generation patterns

This enables legal accountability by linking synthetic media to specific models and their operators.

Combines multiple forensic signals across modalities for robust detection resistant to single-channel attacks:

• Audio-Visual Cross-Verification: Correlates facial movements with speech patterns to detect mismatched modalities
• Text-Image Consistency: Verifies semantic alignment between generated captions and visual content
• Ensemble Decision Logic: Aggregates predictions from physiological, frequency, and temporal detectors with weighted confidence scoring

Fusion approaches defeat targeted countermeasures that may fool individual detection methods.