Federated learning
Definition
Federated learning trains models across many devices or organizations while keeping raw data local. Only model updates (e.g. gradients) are shared, reducing privacy and regulatory risk.
Use it when data cannot be centralized (e.g. hospitals, phones) but you still want a shared machine learning model. Privacy is improved compared to sending raw data; additional techniques (differential privacy, secure aggregation) can be layered. See AI ethics for privacy and governance context.
How it works
The server holds the global model and sends it to clients (devices or orgs). Each client trains locally on its own data and sends updates (gradients or model diff) back. The server aggregates updates (e.g. FedAvg: average the client models or gradients) and produces a new global model, then broadcasts again. Rounds repeat until convergence. Challenges: heterogeneity (non-IID data, different compute), communication cost (limit round count or update size), and privacy (updates can leak information; DP or secure aggregation mitigate).
Use cases
Federated learning fits when data must stay on devices or silos and you still want a shared model.
- Training on sensitive data (e.g. healthcare, finance) without centralizing it
- Mobile and edge devices (e.g. keyboard suggestions, on-device ML)
- Cross-organization collaboration under privacy constraints