How it works
CrashBench generates realistic edits (added damage, water stains, swapped parts) on real photos, then recompresses to mimic trace hiding. The detector trains on exactly the hard case.
Everything is labelled at generation (model, seed, ops) and C2PA-watermarked. The goal is explicit: train the defenses, not bypass them.
Key points
- Coherent fraud corpus to train detectors.
- Recompression to simulate trace hiding.
- C2PA watermark + lawful use required (dual-use guardrail).