Bring the model to the data
In federated learning, raw data never leaves the device. Instead of uploading data to a central server, each client trains the shared model locally on its own data, then sends only the model update back. A server aggregates these updates into a new global model.
The basic round
- The server sends the current global model to a sample of clients.
- Each client computes an update from its local data.
- The server averages the updates, a step often called federated averaging.
- The improved model is sent out again for the next round.
Why it helps
- Raw data stays put, reducing exposure and easing some regulatory constraints.
- It taps data that could never be centralized, such as on phones.
What it does not solve
- Updates can still leak information, so it is often paired with differential privacy or secure aggregation.
- Clients have non identical data distributions, which makes training harder and can slow convergence.
Key idea
Federated learning trains a shared model by sending updates rather than raw data and averaging them on a server, improving privacy and access but still needing extra protection against update leakage and uneven client data.