Federated learning for children's data: encoding trust in education systems
- Venue
- UNICEF Innocenti, Office of Strategy and Evidence
- Year
- 19 May 2025
- Role
- Author
The question
Can public institutions learn from sensitive education data without ever centralising it?
Abstract
Argues that governments can use sensitive education data without centralising it, and what governance that requires.
Role
Author. Published by UNICEF Innocenti as part of its children's data governance project, on whose expert advisory group he also sits.
Authors. Roy Saurabh
Sensitive data should not have to be centralised in order to become useful.
The idea, in four moves
01
The records stay where they are
Each school system keeps its own data. No child's record leaves the boundary of the institution that holds it.
02
Only model updates move
A shared model is trained locally in each place. What travels between institutions is the update, not the data.
03
Governance has to travel with it
Who may join, what may be computed and who answers for the result must be settled before the first update moves.
04
Risk moves; it does not vanish
Updates can leak and explanations can drift. Federation changes where the risk sits. It does not remove it.
Evidence
- 01
UNICEF Innocenti
Federated learning for children's data: encoding trust in education systems
Article authored by Roy Saurabh and published by UNICEF Innocenti on 19 May 2025, listed among the outputs of the good-governance-of-children's-data project.
- 02
UNICEF Innocenti
Good governance of children's data, advisory groups
Names Roy Saurabh among the current expert advisory group members shaping the second phase of UNICEF Innocenti's children's data governance project.
Related
Next project
Assurance
An assurance toolchain for deployed AI