PAD and its limits
Presentation-attack detection (PAD) screens for printed photos and screen replays. It is a signal, and today it is not validated.
No PAD figure has been measured on relevant attack media, so Warka publishes none. Automatic approval stays off until that evidence exists.
What runs today
- MiniFASNet, a single-frame passive classifier, on every usable frame. The most cautious frame wins.
- With live keys, a file identical to an earlier capture in your organization is treated as suspected replay. Sandbox skips this so you can reuse test images, and a changed copy of a file isn't caught.
- With several frames, Warka also checks for no motion and for a changed identity.
What it doesn't cover
- Camera injection, where a modified device feeds in a video.
- Deepfakes and other synthetic media.
- High-quality masks.
- A replayed video with natural motion can pass the motion check.
A pad.score is raw model output, not the probability that someone is real. Warka does not claim ISO/IEC 30107-3 conformance.