Federated Learning
A training approach where models learn from data spread across many devices or organisations without the raw data ever leaving its source — only model updates are shared.
Federated learning trains a shared model across decentralised data. Instead of pooling sensitive records in one place, each participant trains locally on its own data and sends only encrypted model updates to a central server, which aggregates them into an improved global model.
The appeal is privacy and control. Hospitals can collaborate on a diagnostic model without exchanging patient records; phones can improve a keyboard’s predictions without uploading what people type.
Challenges include:
- Communication cost: participants may be on slow or intermittent connections.
- Non-identical data: each participant’s data differs, which can destabilise training.
- Aggregation attacks: updates can leak information about local data unless protected with techniques like secure aggregation or differential privacy.
- Incentives: participants need a reason to contribute compute and data.
Federated learning is closely tied to privacy regulation and on-device AI. It is one of the few techniques that genuinely reduces data concentration risk rather than just promising to.