Media authenticity framework
A structured approach to video verification that goes beyond single-score detection.
12 min read
Why we need a framework
Most detection tools produce a single score: "85% fake" or "92% real." This reductive approach fails to capture the complexity of media authenticity. A video can be technically authentic but contextually misleading. It can have legitimate edits (color grading, cropping) that trigger false positives. It can have provenance data that definitively proves origin regardless of pixel-level analysis.
The ClipForensics media authenticity framework provides a structured methodology for evaluating video authenticity across multiple independent dimensions.
Framework Layers
Five layers of authenticity assessment
Layer 1: Provenance Layer
Does the video have verifiable origin data? C2PA content credentials, IPTC metadata, or blockchain provenance records provide the strongest authenticity signals. When present and valid, provenance data can establish origin with high confidence — independent of any pixel-level analysis.
Layer 2: Container Layer
What does the file structure reveal? Metadata consistency, compression history, encoding chain, and codec signatures provide context about how the video was created and processed — before examining any visual content.
Layer 3: Content Layer
What do the pixels and audio reveal? Visual artifact detection, spectral analysis, face manipulation, lip-sync alignment, audio synthesis detection, and temporal consistency form the core forensic analysis.
Layer 4: Fusion Layer
How do all signals combine? Weighted evidence fusion with agreement bonuses and contradiction penalties produces a single trust score with a confidence interval. Cross-module correlation analysis identifies consistent vs. conflicting signals.
Layer 5: Reporting Layer
How are results communicated? Clear verdict categories, per-module breakdowns, evidence timelines, and honest confidence levels. Results should enable informed decision-making, not dictate conclusions.
Framework principles
Multi-dimensional over single-score
Authenticity is not binary. A comprehensive assessment examines multiple independent dimensions and reports findings transparently, rather than collapsing everything into one number.
Evidence over claims
Every conclusion must be backed by specific evidence. Reports should show which modules contributed, what they found, and how confident each assessment is.
Uncertainty over false confidence
When evidence is insufficient or contradictory, the correct response is "Inconclusive" — not a guess. False confidence is more dangerous than acknowledged uncertainty.
Limitations transparency
Every detection approach has blind spots. A credible framework explicitly states its limitations rather than implying universal capability.
Provenance first
When cryptographic provenance data (C2PA) is available and valid, it provides stronger evidence than any forensic analysis. The framework prioritizes provenance verification when available.
Applying the framework
ClipForensics implements this framework through its 15 forensic modules, evidence timeline, and compression history features. Each module maps to specific framework layers, and the fusion engine combines them according to the principles described above.
We encourage other researchers and organizations to adopt, adapt, or critique this framework. Media authenticity is a collective challenge that benefits from shared methodological standards.