15 forensic analysis modules
Each module examines a different dimension of the video. Together, they provide a comprehensive authenticity assessment.
Detection Engine
Module reference
Every module runs independently and produces its own confidence score and evidence list.
| Module | Category | Weight | Description |
|---|---|---|---|
| Metadata Analysis | Container | 10% | Inspects encoder tags, creation timestamps, software signatures, and container-level metadata for inconsistencies. |
| Compression Forensics | Container | 10% | Analyzes codec signatures, bitrate patterns, GOP structures, and quantization parameters for signs of re-encoding. |
| Compression History | Container | 5% | Reconstructs the full encoding chain to determine how many times a video has been compressed and by which tools. |
| Provenance Verification | Container | 10% | Checks for C2PA content credentials, IPTC data, and other provenance signals that verify origin. |
| Visual Artifact Detection | Visual | 8% | Error Level Analysis (ELA), noise pattern analysis, and GAN checkerboard detection at the pixel level. |
| Temporal Consistency | Temporal | 8% | Frame-to-frame coherence analysis — checks for sudden resolution changes, temporal aliasing, and splicing artifacts. |
| Face Manipulation Detection | Visual | 10% | Detects face-swap artifacts, facial boundary inconsistencies, and deepfake-specific anomalies around eyes, teeth, and hairline. |
| Lip Sync Analysis | Audio-Visual | 5% | Measures phoneme-viseme alignment to detect lip-sync puppetry and audio-driven face animation. |
| Audio Synthesis Detection | Audio | 5% | Identifies text-to-speech artifacts, voice cloning signatures, and spectral anomalies in the audio track. |
| Audio-Visual Sync | Audio-Visual | 5% | Cross-modal timing verification — detects desynchronization between audio events and visual events. |
| Optical Flow Analysis | Temporal | 5% | Evaluates motion field consistency, checks for physically implausible movements, and detects warping artifacts. |
| Spectral Analysis | Visual | 5% | Frequency-domain analysis using DCT and FFT to detect generative model signatures invisible to the eye. |
| Biological Motion | Temporal | 5% | Assesses naturalness of human movement patterns including gait, gesture timing, and micro-expressions. |
| Lighting Consistency | Visual | 4% | Analyzes illumination direction, shadow consistency, and specular highlights for composite detection. |
| Localized Manipulation | Visual | 5% | Region-specific analysis that maps potential manipulation areas within individual frames. |
Module Categories
Four analysis dimensions
Container
File-level analysis — metadata, compression signatures, encoding history, and provenance credentials. These modules examine the video before looking at any pixels.
Visual
Pixel-level analysis — artifact detection, spectral frequency, lighting consistency, face manipulation, and localized region mapping.
Temporal
Motion and time analysis — frame-to-frame consistency, optical flow physics, and biological motion naturalness.
Audio / Audio-Visual
Sound analysis and cross-modal verification — voice synthesis detection, lip-sync alignment, and audio-visual synchronization.
Evidence Fusion
How module scores become a verdict
Module results are fused using calibrated weights, agreement analysis, and contradiction detection.
Each module produces a confidence-weighted score between 0 and 1. These scores are combined using pre-calibrated weights (shown in the table above).
The fusion engine applies agreement bonuses when multiple independent modules reach similar conclusions, and contradiction penalties when modules disagree — forcing the final score toward "Inconclusive" rather than guessing.
The resulting trust score maps to one of six verdict categories, each with a clearly defined confidence threshold. Read more about our scoring methodology.