Detection Methodology

1Our Approach

VeriVid AI uses an ensemble pipeline that combines multiple detection signals to estimate the likelihood that a video is AI-generated or manipulated. We analyze visual, audio, and metadata characteristics to produce a risk score.

2Visual Analysis

We extract key frames from the video and analyze them for synthetic artifacts:

• ✓AI Generation Patterns: Detection of artifacts common in AI-generated imagery (GAN fingerprints, diffusion patterns)
• ✓Facial Analysis: When faces are present, we check for deepfake indicators and manipulation signs
• ✓Consistency Checks: Temporal consistency across frames to detect splicing or injection

3Audio Analysis

For videos with audio, we analyze the speech and sound characteristics:

• ✓TTS Detection: Identification of text-to-speech synthesis patterns
• ✓Voice Cloning Artifacts: Analysis for AI voice cloning signatures
• ✓Audio-Visual Sync: Checking lip sync and audio alignment

4Metadata & Tampering

We examine file metadata for manipulation indicators:

• ✓Encoding Analysis: Multiple re-encoding, suspicious codec combinations
• ✓EXIF/Metadata: Missing camera info, inconsistent timestamps
• ✓Container Analysis: File structure anomalies

5Risk Scoring

Individual signals are weighted and combined into a final risk score:

!Known Limitations