Face Recognition
Identifying a person from their face — technically solved, and the single clearest case of a system that works well on average and fails on specific people.
When not to use it
- 1:N identification with a large gallery. False positives scale with N. The quoted accuracy is the 1:1 number.
- Where a false positive means arrest. The error distribution falls hardest on people already over-policed, and it has happened.
- Without disaggregated evaluation. An aggregate rate averaged over a population that doesn't experience the system equally describes nobody.
- On anyone who didn't consent. For verification, consent is inherent. For surveillance it's impossible.
Reach for something else instead
- Other biometrics — fingerprint, iris. Require cooperation, which is the point.
- Non-biometric identity — badges, cards, passwords. Revocable, which faces aren't.
- Human review as a gate — with the rubber-stamp caveat.
- Not identifying people — a lot of stated use cases don't actually require knowing who someone is.
Sources & further reading
- Buolamwini & Gebru (2018), Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification — 0.8% vs 34.7%. The paper that changed the field.
- Grother, Ngan & Hanaoka (2019), NIST FRVT Part 3: Demographic Effects — 189 algorithms, 99 developers. The independent confirmation at scale.
- Schroff, Kalenichenko & Philbin (2015), FaceNet: A Unified Embedding for Face Recognition and Clustering — the architecture everything still uses.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Quoting a verification accuracy number in an argument about surveillance. Different problem, different error profile.
- Treating one accuracy number as the accuracy. Gender Shades is exactly what that conceals.
- Assuming more data fixes it entirely. It helps a lot; the single-threshold problem remains.
- Forgetting your face isn't revocable. A leaked password can be changed.
At a glance
FieldComputer Vision
The finding0.8% error for lighter men, 34.7% for darker women (Gender Shades, 2018)
ConfirmedNIST, 189 algorithms, 99 developers
Two problems1:1 verification (safe) vs 1:N identification (false positives scale with N)
The unfixable bitno single threshold is fair to every group
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Verification vs. identification — one asks "are you who you claim" with your consent and one comparison; the other asks "who are you" against a million strangers. Only the first is nearly solved.