Book a briefing
Check 04 · face match

Is it the same
face?

The face match compares the member’s live selfie with the photo on their identity document and returns a similarity score. A clear match passes the check; a borderline score goes to a reviewer; a clear mismatch fails it.

512
numbers in the face template the model produces for each photo
2
photos compared: the live selfie and the document photo
3
outcomes: passes, goes to a reviewer, or fails
None
accuracy figure published — we have not measured one we would stand behind
What the software does

Find, describe,
compare.

Four steps, run on PollGuard’s verification service each time a member submits a verification.

  1. 01

    Find the face

    A standard face detector (OpenCV’s ResNet-SSD model) finds the face in the selfie and in the document photo, and crops it.

    Detection
  2. 02

    Describe it

    An ArcFace ResNet-100 model, a widely used open face-recognition model, turns each cropped face into a template of 512 numbers.

    Face template
  3. 03

    Compare

    The two templates are compared with cosine similarity: the closer the score is to 1, the more alike the two faces look to the model.

    Similarity score
  4. 04

    Decide the check

    The score is set against fixed thresholds. Anything in the uncertain band, or any photo where no face was found, waits for a person.

    Pass · review · fail
The numbers, as set today

Thresholds,
not accuracy.

These are the cut-offs the pipeline uses to sort a score. They say what happens to a score; they do not say how often the model is right. That we have not measured.

Face match settings

0.42 and above
Check passes.The member still waits for a reviewer’s approval.
0.30 – 0.42
Goes to a reviewer.The score is shown to them with both photos.
No face found
Goes to a reviewer.A missing face is never read as a match or a mismatch.
Below 0.30
Check fails.Recorded with the score, so the decision can be explained.

Read from the verification pipeline on 17 Oct 2026 (set 13 Oct 2026). They can change; this page will change with them.

What is kept

The selfie’s face template is stored with the verification, alongside the photos and scores the reviewer needs. It is a list of numbers, not a picture, but it is still personal data and is treated as such.

Where it runs

On PollGuard’s verification service, not on the phone. The phone captures the photos; the comparison happens on the server.

The risk we still carry

Face-recognition models can perform differently across groups of people. We have not yet measured how this one performs for PollGuard’s members. Until we have, a person reviews every case.

A person approves every member.
No check on this page verifies anyone by itself. The results go to a reviewer for the organisation, who approves or rejects — and that decision, with their name and any note they add, is written to the record. Where a member needs help at a station, a returning officer assists and is named too.
What it does not do

What the face match
does not do.

Give an accuracy percentage
We publish none, and we will not until we have measured one on real conditions and can show the working.
Identify a stranger
It compares two photos the member supplied. It does not search for faces across the internet or any outside database.
Check a government register
There is no lookup against IPRS or any government database.
Approve a member
A passing score is evidence for the reviewer, not a decision.
The four checks

Ask us how it
handles your roll.

Bring your questions to a briefing — about review queues, officer assists, data protection and what happens when a check is unsure.

Book a briefing Read the trust page