The mathematics of anti-spoofing, without the hype

This piece is published as an outline: the statement and the structure are settled, the full text is coming.

The problem, stated

Given an image, decide whether there is a person in front of the camera or a reproduction. A binary classifier with deeply asymmetric costs.

Two distributions that overlap

Genuine and impostor scores are not two separate clouds. The whole design is about where you cut the overlap.

The ROC curve and the point you pick

Why the optimal threshold depends on what each error costs, and in attendance control those costs look nothing alike.

What the real world measures

Lighting, cheap cameras and users in a hurry: why a lab threshold never survives the first shift.