Deepfake Detection

Transforms an image from its spatial representation into the frequency domain using Discrete Fourier Transform (DFT) or Discrete Cosine Transform (DCT). This exposes grid-like artifacts and unnatural high-frequency patterns generated by neural network upsampling layers that are invisible to the naked eye. Generative models often leave periodic peaks in the power spectrum due to transposed convolution operations. Analysts look for spectral centroids and banding anomalies that deviate from the 1/f law typical of natural photography.

A liveness detection method that extracts subtle, periodic variations in skin color caused by blood volume changes during the cardiac cycle. By magnifying these micro-color changes in facial regions of interest and analyzing the frequency spectrum, detectors can verify the presence of a living human cardiovascular system. Deepfake faces often fail to replicate the coherent spatiotemporal dynamics of genuine blood flow, exhibiting flat or non-physiological pulse signals.

A specific audio-visual forensic analysis that detects inconsistencies between spoken phonetic sounds and the observed mouth shapes required to produce them. A phoneme (the distinct unit of sound) must correspond to a specific viseme (the visual mouth configuration). Deepfake generators often produce plausible but incorrect viseme sequences, especially for bilabial consonants like /p/, /b/, and /m/, or fricatives like /f/ and /v/, where lip closure or teeth contact is physically precise.

The process of identifying unique, inherent artifacts left in synthetic images by the specific Generative Adversarial Network architecture used to create them. Every generator leaves a systematic 'fingerprint' in the pixel space, often a consequence of its specific upsampling operations, normalization layers, or training pipeline. These fingerprints can be extracted using a dedicated classifier trained on the noise residuals of known model families, enabling attribution to architectures like StyleGAN2, ProGAN, or StyleGAN3.

A source camera identification method that extracts the unique, stable sensor pattern noise caused by silicon manufacturing imperfections. This Photo Response Non-Uniformity (PRNU) acts as a ballistic fingerprint for a specific camera sensor. By correlating the extracted noise residual from a query image against a known camera's reference pattern, analysts can verify if an image originated from that specific device. A missing or corrupted PRNU pattern suggests the image has been synthetically generated or heavily manipulated.

A forensic technique that fits a parametric three-dimensional face model to a two-dimensional image to detect inconsistencies in the estimated shape, texture, and lighting parameters. By solving for the 3D geometry, albedo, and illumination coefficients that best explain the image, the system can flag faces where the recovered parameters are physically implausible. Face-swapped deepfakes often produce mismatches between the estimated 3D shape of the face and the surrounding head geometry or scene lighting.

A core deepfake detection metric that measures the temporal and spatial misalignment between a subject's visual lip movements and the corresponding audio speech track. Early face-swapping models often fail to perfectly synchronize mouth shapes with phonemes.

• Analyzes phoneme-viseme correspondence
• Detects frame-level audio-visual lag
• Flags unnatural jaw and lip transitions
• Often combined with MFCC forensics for robust scoring

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. Synthetic faces generated by GANs and diffusion models typically lack these microscopic, rhythmic color fluctuations.

• Measures remote heart rate signals from facial regions
• Analyzes chrominance changes in the RGB color space
• Detects absence of physiological signal as a spoof indicator
• Effective against both deepfakes and replay attacks

A forensic technique that transforms an image into its frequency representation using Discrete Fourier Transform (DFT) or Discrete Cosine Transform (DCT) to detect anomalies invisible in the spatial domain. AI-generated faces often exhibit unnatural high-frequency patterns from neural network upsampling layers.

• Reveals grid-like artifacts from transposed convolutions
• Identifies spectral peaks absent in optical camera capture
• Detects GAN fingerprints in the Fourier magnitude spectrum
• Complements spatial-domain classifiers for ensemble detection

A comprehensive anatomical system for coding individual facial muscle movements called Action Units (AUs). Used as a ground-truth framework to detect unnatural or impossible AU combinations in deepfakes that violate human facial anatomy constraints.

• Defines 46 unique Action Units for muscle groups
• Flags AU co-occurrences that are anatomically impossible
• Measures AU activation timing and symmetry
• Deepfake models often produce bilaterally symmetric but unnatural expressions

A forgery detection method that compares the position, shape, and intensity of specular reflections in the eyes or on surfaces. In composite images where a face is spliced from a different source, the corneal reflections often reveal inconsistent light source directions.

• Analyzes corneal specular highlights in both eyes
• Estimates 3D light direction from reflection geometry
• Detects physically impossible lighting configurations
• Effective against face-swap and reenactment deepfakes

The classification of speech audio as genuine or synthetically generated by analyzing artifacts in the spectral domain, prosody, and vocoder fingerprints. Modern neural vocoders leave characteristic traces in high-frequency bands that differ from human vocal tract acoustics.

• Uses Mel-Frequency Cepstral Coefficients (MFCC) as features
• Detects vocoder-specific artifacts above 8 kHz
• Analyzes unnatural pitch contour and rhythm patterns
• Often deployed alongside visual detection for multimodal verification