Video Forensics and AI Detection Analysis
Hypothesis
Can non-AI mathematical analysis (pixel-level, signal processing) provide more robust video forensics than AI-based detection, especially against adversarial attacks?
Approach
- Analyzed AI-powered detection tools (Copyleaks, Hive, Attestiv)
- Researched non-AI mathematical approaches (ELA, PRNU, Fourier)
- Designed “dirty data” testing pipeline
- Developed co-training architecture for hybrid detection
Findings
Key Observations
AI Detection Limitations:
- “Black box” solutions lack explainability
- Generation models (GANs/Diffusion) eventually outpace detection
- Susceptible to adversarial attacks
Non-AI Mathematical Approaches:
| Method | What It Detects | How It Works |
|---|---|---|
| Error Level Analysis (ELA) | Compression anomalies | Re-compress and subtract |
| Sensor Pattern Noise (PRNU) | Camera fingerprint mismatch | Silicon imperfection signatures |
| Fourier Transform | High-frequency artifacts | Spatial to frequency domain |
| Benford’s Law | Normalized pixel distributions | First digit statistics |
Dirty Data Pipeline Architecture:
Stage 1: Clean Generator (Remotion/SORA)
↓
Stage 2: Corruption Layer (noise, compression, frame drops)
↓
Stage 3: Hybrid Analysis Gauntlet
├── AI Track (Copyleaks/Hive)
└── Non-AI Track (ELA/PRNU)
↓
Stage 4: Discrepancy Evaluation
RSI (Recursive Synthetic Improvement) Pipeline:
Integration with Remotion/ReVideo/Kdenlive for co-training:
- Non-AI (Math): Rigid Verifier providing ground-truth constraints
- Local AI (Jan): Pattern Intuitionist learning visual patterns
- Feedback Loop: Each trains the other iteratively
Co-Training Protocol:
- Remotion generates 1000 video clips (clean/manipulated/dirty)
- Non-AI tools confidently label 60% based on physics
- Local AI trains on confirmed labels
- Local AI attempts to classify remaining 40%
- New discoveries feed back to improve Non-AI parameters
Outcome
Status: Success
Hybrid approach (Non-AI + AI) provides robust detection with explainability.
Next Steps
- Implement corruption layer script (FFmpeg/Python)
- Create Remotion adversarial video generator
- Set up co-training pipeline with Jan
- Benchmark detection accuracy on dirty data
Related
- Remotion - Video generation
- Kdenlive-Blender - Production pipeline
- Optimal Small AI - Local AI training