Vision foundation models fail at open-set iris spoofing detection, study finds
Even top models can't spot unknown iris attacks reliably.
Deep Dive
Researchers conducted a systematic failure analysis of five vision foundation models for iris Presentation Attack Detection (PAD) under realistic open-set conditions. Using three protocols—unseen attack instruments, unseen datasets with different sensors, and cross-spectral (NIR to VIS) shifts—they found models transfer well across similar datasets but fail on unseen attacks and degrade sharply across spectra. LoRA fine-tuning sometimes worsens these failures.
Key Points
- Five vision foundation models tested under three open-set protocols (unseen attack instruments, unseen datasets, cross-spectral NIR→VIS).
- Models fail to generalize to unseen attacks and degrade sharply under cross-spectral shifts.
- LoRA fine-tuning often makes failures worse under attack-level and spectral shifts, not better.
Why It Matters
Iris liveness detection systems relying on vision models may be vulnerable to novel spoofs, demanding new robust PAD approaches.