How it works
Insurers are drowning in retouched or inpainted claim photos. To detect fraud you need fraud examples — coherent, varied, labelled.
The deterministic stage then recompresses the image to hide editing traces, exactly as a fraudster would. Your detector trains on the hard case.
Key points
- Realistic prompt-based editing for detector training data.
- C2PA watermark and lawful-use clause (dual-use guardrail).
- Recompression to simulate hiding editing traces.