Generative Document Forgery with RL
Using reinforcement learning to steer image-generation foundation models toward realistic document forgeries for robustness testing. In progress.
Status: in progress, with Prof. Andreas Dengel and Dr. Sheraz Ahmed (DFKI).
Our BMVC 2026 paper showed that tampering detectors can be broken by forgeries built to exploit what they rely on. This project scales that idea up. Instead of hand-designing attacks, we use reinforcement learning to fine-tune image-generation foundation models, rewarding them for forgeries that are realistic, correctly localized, and challenging for detectors. The generated data then serves two uses:
- training data for more robust tampering detectors, and
- stress tests that reveal detector failure modes before real attackers find them.
Copy-move, splicing and generative forgeries, from our BMVC 2026 paper. This project learns to generate the generative case.
As groundwork for text editing with diffusion models, I am reproducing DiffUTE from its paper: DiffUTE-repro.