Video Verify

AI-powered video authenticity checker for deepfake detection using GPT-4 Vision.

Features

• Video Upload: Drag-and-drop or click to upload video files (MP4, MOV, WebM)
• Multi-Segment Sampling: Analyzes 24 frames across 3 segments (25%, 50%, 75% of video)
• Structured Scoring: Per-region analysis with 0-10 scores (mouth, eyes, boundary, temporal)
• Degree of Risk: Clear 0-10 risk score with color-coded visualization
• Demo Mode: Test with pre-loaded Morgan Freeman deepfake video
• Report Download: Export detailed analysis results

How It Works

1. Upload a video file (or click "Test with Morgan Freeman Deepfake")
2. System extracts 24 frames across 3 video segments (8 frames each at 25%, 50%, 75%)
3. Frames are sent to GPT-4 Vision with structured scoring prompt
4. AI analyzes each region and returns 0-10 scores:
   • Mouth: Teeth clarity, lip texture, movement naturalness
   • Eyes: Reflection matching, pupil consistency, blink patterns
   • Face Boundary: Color matching, jawline blending, hairline artifacts
   • Temporal: Frame-to-frame consistency, flickering, warping
5. Results displayed with per-region scores and overall Degree of Risk
6. Download report for documentation

Tech Stack

• Frontend: Next.js 16, React 19
• AI: OpenAI GPT-4 Vision with JSON response format
• Hosting: Vercel
• Frame Extraction: Client-side Canvas API
• Video Processing: Multi-segment temporal sampling

Local Development

# Install dependencies
npm install

# Set environment variable
echo "OPENAI_API_KEY=your-key-here" > .env.local

# Run development server
npm run dev

Deployment

Deploys automatically to Vercel on push.

Live demo: https://video-verify.cameronobrien.dev

Detection Accuracy

Based on research, GPT-4 Vision achieves 77-79% AUC for deepfake detection among vision-language models. The structured scoring approach with per-region analysis improves reliability by forcing numerical commitments rather than hedged prose responses.

Limitations

• Analysis quality depends on video resolution
• Works best with videos containing faces
• Not a definitive forensic tool (use for screening)
• VLM-based detection has inherent accuracy limits vs specialized CNN models

License

MIT