Face ID and Face Recognition

Apple introduced Face ID with the iPhone X in 2017. It is the most familiar example of face recognition, though it solves a narrower problem: verification (“is this the phone’s owner?”) rather than identification (“who in this crowd is this?”).

How it works:

1. A dot projector casts thousands of invisible infrared dots on your face; an infrared camera reads them to build a depth map.
2. On-device neural networks turn the depth map and infrared image into a compact numeric representation, an embedding.
3. That embedding is compared with the one stored at enrollment, inside a secure chip. Close enough: unlocked.
4. Separate networks check for spoofing attempts such as photos or masks.

Everything runs on the phone (edge AI) and the template never leaves the device. Apple says the chance a random person could unlock your phone is below 1 in a million, though higher for twins, look-alike siblings and young children.

The same embedding idea powers general face recognition, which is far more controversial when used for surveillance: it raises privacy questions, and NIST testing found that many algorithms had higher error rates for some demographic groups. The EU AI Act restricts real-time remote biometric identification in public spaces. See also bias and fairness.