On this page
Short answer. Liveness detection answers "is a real person present?". Face matching answers "is it the person on this ID?". Run liveness first, then match, then decide. Each one fails in ways the other cannot catch.
Two questions, two failure modes
People often use the two terms as if they were one feature. They are separate checks with separate ways to go wrong.
| Liveness | Face matching | |
|---|---|---|
| Question | Is a live person present right now? | Is this the same person as on the document? |
| Input | The selfie or video capture | The selfie and the ID portrait |
| Stops | Photos, replays, masks | Someone using another person's genuine ID |
| Misses | A live person with a stolen ID | A spoofed face of the right person |
A criminal with a stolen passport and their own live face passes liveness and fails the match. A criminal with a photo of the right person passes the match and fails liveness. You need both to close both doors.
The order that saves money and risk
- Quality gate. Reject blurred, dark or occluded images with a retry prompt.
- Liveness. If this fails, stop. There is no point matching a spoof.
- Document read. Extract and sanity-check the fields on the ID.
- Face match. Compare the live face with the document portrait.
- Decision. Accept, send to review, or reject, using all the scores together.
Failing early is cheaper in compute and safer: a bad actor learns nothing about your match threshold if they never get past liveness.
Understanding the match score
A face matcher turns each face into a vector of numbers and measures how close they are. The result is a similarity score. It is not a probability, and scores from different vendors are not comparable. What matters is where you place your threshold:
- Higher threshold: fewer wrong people accepted, more genuine users sent away.
- Lower threshold: smoother onboarding, more risk of accepting a look-alike.
For IDNexus Standard we suggest a similarity of 0.68 or higher as a starting point. Treat it as a default to tune on your own traffic, not a law of nature.
Use a middle band. Instead of one cut-off, define three zones: clearly above your threshold, accept; clearly below, reject; in between, manual review. The middle band is where most disputed cases live, and a human look there is cheap insurance.
Why ID photos make matching harder
The portrait on an ID is often years old, small, compressed and printed on glossy stock. Glasses, beards, weight change and lighting all pull the score down. This is why a face matcher should be paired with a review lane and clear retry guidance (face the light, remove glasses) rather than a hard binary.
What IDNexus returns
Standard and Enterprise responses include the liveness result, the document read and a face-match similarity score, so you can apply your own rules. The score and model version are part of the result, which makes each decision explainable later. See the API documentation for the request format.
See both scores in one response
A Standard verification returns liveness, document and face-match results together for 50 credits ($0.50).
Create account