Learn

Video forensics glossary

Key terms and definitions for video forensics, deepfake detection, and media authenticity.

A

Artifact: An unintended visual or auditory anomaly introduced during generation, compression, or manipulation. Forensic analysis detects artifacts that are characteristic of specific manipulation techniques.
Autoencoder: A neural network architecture that compresses input data into a lower-dimensional representation and reconstructs it. Used in face-swap deepfakes with shared encoder weights between two identities.

Biological motion: The characteristic patterns of human movement governed by biomechanical constraints. AI-generated humans often exhibit subtle deviations from natural biological motion.
Bitrate: The amount of data used to represent one second of video, typically measured in Mbps. Variable bitrate (VBR) allocates more data to complex scenes; constant bitrate (CBR) maintains steady data rate.

C2PA: Coalition for Content Provenance and Authenticity. An industry standard for embedding cryptographically signed provenance metadata in media files, establishing a chain of trust from capture to publication.
Codec: Compressor-decompressor. Software that encodes video for storage/transmission and decodes it for playback. Common video codecs: H.264, H.265/HEVC, VP9, AV1.
Compression history: The sequence of encoding operations a video has undergone. Each re-compression leaves forensic traces that can be detected through quantization analysis and DCT coefficient distributions.
Confidence interval: A range that quantifies the uncertainty of a forensic assessment. ClipForensics reports confidence levels for each module and the overall verdict.
Content credentials: Cryptographic metadata embedded in media files that verify the origin, creation tool, and editing history. Based on C2PA standards.

DCT: Discrete Cosine Transform. A mathematical operation central to JPEG and video compression. Transforms spatial data into frequency coefficients. DCT coefficient analysis is a key forensic technique.
Deepfake: Synthetic media created or modified using deep learning. Originally referred to face-swap videos; now encompasses any AI-generated or AI-manipulated media including audio.
Diffusion model: A class of generative AI models that create data by learning to reverse a noise-adding process. Used by Sora, Stable Diffusion, and other modern generators.

ELA: Error Level Analysis. A forensic technique that re-compresses an image at a known quality level and analyzes the error distribution. Manipulated regions often show different error patterns.
Evidence fusion: The process of combining results from multiple independent forensic modules into a single assessment using weighted scoring, agreement bonuses, and contradiction penalties.
Evidence timeline: A chronological mapping of forensic findings to specific timestamps in a video, showing when and where anomalies were detected.

Face reenactment: Manipulating a person's facial expressions to match a driving video. The identity remains the same but the expressions are puppeteered.
Face swap: Replacing one person's face with another in video while preserving head pose, expression, and lighting. One of the most common deepfake techniques.
FFT: Fast Fourier Transform. Converts signals from the time/spatial domain to the frequency domain. Used in spectral analysis to detect generation artifacts invisible to the human eye.
Forensic module: An independent analysis component that examines one dimension of a video (e.g., metadata, compression, face manipulation). ClipForensics uses 15 independent modules.

GAN: Generative Adversarial Network. A neural network architecture with a generator-discriminator pair. The generator creates synthetic data while the discriminator evaluates its realism.
GOP: Group of Pictures. A sequence of video frames between keyframes. GOP structure and size can reveal the encoder used and whether the video has been re-encoded.

Lip sync: The alignment between mouth movements (visemes) and speech sounds (phonemes). Lip-sync manipulation modifies mouth movements to match new audio.
Liar's dividend: The phenomenon where the existence of deepfake technology allows people to dismiss authentic footage as fake. Even without creating deepfakes, the technology undermines trust in genuine media.

Metadata: Data about data. In video forensics: encoder tags, creation timestamps, GPS coordinates, camera model, software identifiers, and other information embedded in the file container.

Noise pattern: The characteristic noise produced by a camera sensor (PRNU — Photo Response Non-Uniformity). Unique to each sensor. AI-generated content has fundamentally different noise characteristics.

Optical flow: The apparent motion of objects between consecutive video frames. Computed by analyzing pixel displacement. Used to detect physically implausible motion in manipulated videos.

Perceptual fingerprint: A compact numerical representation of a video's visual content that is robust to re-encoding, cropping, and quality changes. Used for cross-video matching.
Phoneme: The smallest unit of speech sound. Forensic lip-sync analysis compares phoneme timing with corresponding mouth shapes (visemes) to detect manipulation.
Provenance: The documented origin and processing history of a piece of media. Strong provenance data (C2PA, IPTC) provides the most reliable authenticity signal.

Quantization: The process of reducing the precision of values during compression. Quantization tables and parameters are encoder-specific forensic signatures.

Spectral analysis: Analysis in the frequency domain rather than spatial domain. Reveals patterns invisible to the eye, including GAN checkerboard artifacts and diffusion model signatures.

Trust score: ClipForensics's overall authenticity assessment, ranging from 0 (high manipulation risk) to 1 (high authenticity confidence). Based on weighted fusion of all module scores.

Verdict: The final classification assigned to an analyzed video. ClipForensics uses six categories: Verified Origin, Likely Authentic, Likely AI-Edited, Likely AI-Generated, High Manipulation Risk, and Inconclusive.
Viseme: The visual representation of a phoneme — the mouth shape corresponding to a speech sound. Used in lip-sync forensic analysis.
Voice cloning: Creating synthetic speech that mimics a specific person's voice using AI. Modern systems can clone a voice from minutes of sample audio.