Facial Recognition

In one line

It is being used in India, and no statute says it may be.

What the technique does

Detection. The software finds a face in an image or a video frame.

Normalisation. It corrects for angle, scale and lighting, which is where most of the error is introduced.

Feature extraction. It reduces the face to a numerical vector, a set of measurements of the geometry of the face which the software treats as a template. The template, not the photograph, is what is compared.

Comparison. The template is compared with one other template, or with every template in a gallery, and the system returns a similarity score.

Threshold. A human or a rule decides what score counts as a match. Nothing in the technology fixes it. The threshold is a policy choice, and it decides the whole error profile of the system.

Verification and identification, which is the distinction that matters

Almost every criticism of facial recognition is a criticism of identification, not verification. A system that is highly accurate one to one can be poor one to many, and quoting an accuracy figure without saying which mode it refers to is the commonest mistake made about this technology.

Why the errors are not random

Image quality. The reference photographs in a police database are taken in controlled light. The probe image from a street camera is at an angle, in poor light, at distance, often of a moving person. The mismatch between the two is the largest single source of error.

Demographic differentials. Error rates are not the same across groups. Reported differentials by skin tone, sex and age mean that the burden of a false match does not fall evenly, which turns a technical defect into a question of equality before the law.

Base rates. In a search of a large population for a rare target, even a small false positive rate produces far more false alarms than true hits, because the number of innocent people is enormous and the number of wanted people is small. This is arithmetic, not a defect of any particular product, and it is the reason a watchlist system generates work rather than arrests.

Automation bias. An officer shown a ranked list tends to accept the top candidate. The system was meant to produce a lead and in practice produces a conclusion.