Diffusion-Based Authentication of Copy Detection Patterns

Counterfeit detection systems based on simple similarity metrics are vulnerable to modern generative attacks. This project introduces a diffusion-based authentication framework that jointly conditions on the binary template, printed CDP, and printer identity to classify source-printer signatures and detect inconsistencies.

Key outcomes:

• Reformulated CDP authentication as per-printer classification with class-conditioned noise prediction.
• Achieved low balanced error on Indigo 1x1 Base (Perr = 0.023).
• In known-versus-unseen counterfeit evaluation, obtained Pfa = 0.000 with Perr = 0.012.

Tech stack:

• Python
• PyTorch
• Stable Diffusion
• ControlNet

Links:

• Project page
• Paper PDF