Facial Recognition At The School Gate
Green Valley School proposes replacing roll call with a camera at the gate that recognises students' faces and marks attendance automatically. The vendor's demo on 40 staff members showed 97% accuracy. The school has 900 students, many of whom arrive in low morning light, wear helmets or dupattas, or are between 11 and 15 years old, an age group the vendor's training data barely covers. The principal likes the time saved, about 12 minutes per class per day. A parent asks what happens to the face images and who can see them. The school has no written policy, and the proposed system would stream video to a cloud service hosted outside the country.
1. Assess the reliability of the vendor's 97% figure for this school, using base-rate reasoning where it helps.
Read the worked answer
The 97% was measured on 40 cooperative adults under the vendor's conditions, so it estimates performance for a different population in different light. Children's faces change shape across 11 to 15, and helmets, dupattas, glasses and low sun all occlude the region a matcher relies on. Expect the real error rate to be worse, and likely worse for some groups than others, which is a fairness problem as well as an accuracy one.
Then apply base rates. A gate camera makes two kinds of error: false accepts, where the wrong student is marked present, and false rejects, where a student standing there is marked absent. At a 3% error rate over 900 students, roughly 27 students would be mis-marked each day. That is far more individual harm than 12 minutes per class of staff time saved. The saving is real but it accrues to staff, while the cost falls on students who must appeal an absence. A demo of 40 adults contains none of these hard cases.
2. Identify the ethical and data-protection issues the school must resolve, and explain why they are serious for this particular group of users.
Read the worked answer
Students are minors, and biometric data cannot be reissued: you cannot give a child a new face after a leak. In India, the DPDP Act 2023 treats the data of under-18s as children's data, requiring verifiable parental consent and barring tracking and behavioural monitoring. A school should therefore assume it needs a lawful basis, clear notice, a defined purpose and a way to withdraw.
Concrete issues: no written policy exists, so nobody can say what is stored or for how long; streaming to an overseas cloud transfers children's biometrics beyond the school's control; and there is no appeal route for a student wrongly marked absent. The minimum fixes are collecting a template rather than photographs, processing on the school server, a retention period such as deletion at year end, access limited to named staff, and an equivalent manual option for any student who opts out.
3. Recommend a decision to the principal. State whether the system should be deployed and under what conditions, with an alternative if it should not.
Read the worked answer
Do not deploy gate-wide this term. Recommend a bounded decision instead.
- Run a four-week pilot with written parental consent, with manual roll call running in parallel so no student is penalised by a camera error.
- Keep processing on-site, store templates only, delete them after the pilot, and publish both the policy and the results.
- Measure false rejects and false accepts separately, broken down by class, lighting and whether the student wore head covering, against a threshold agreed in advance.
- Compare against a cheaper, less intrusive option: an RFID or QR card, or a teacher's tablet roll call, which solves the same 12-minute problem without biometrics.
If the pilot shows unequal error rates between groups, stop and take the card system. The recommendation is conditional rather than a blanket refusal: the school gets its time saving only if the error burden is fairly distributed and families have a real choice.