What is federated learning?
Federated learning is a way of training a machine learning model across several organizations, sites, or devices without moving their raw data to a central place. Each participant trains on its own data locally and sends only model updates to a coordinator, which combines them into a shared model. The data never leaves the hospital, the branch, or the factory where it was created.
How federated learning works
The coordinator starts with an initial model and sends it to every participant. Each participant trains for a few passes on its local data and sends back the changes to the model weights. The coordinator averages those updates, weighted by how much data each participant has, and sends the improved model out again. The cycle repeats until the model stops improving. Cross-silo federated learning involves a handful of organizations with large datasets, such as hospitals or banks. Cross-device federated learning involves millions of phones or sensors, each holding a little data.
Updates can still leak information about the data that produced them, so serious deployments add secure aggregation, which prevents the coordinator from seeing any single participant’s update, and differential privacy, which adds noise to bound what any update reveals.
Federated learning in practice
An industrial equipment maker wants a failure prediction model for compressors installed at customer plants. The customers refuse to send production data off site. The maker places a training agent on the edge gateway at each plant, and the gateways train locally and return weight updates on a weekly cycle. The resulting model predicts failures better than any single site could train alone, and every customer’s process data stays behind its own firewall. The contracts spell out who owns the trained model and how it may be used.
What to watch for
- Participants’ data is rarely alike, and a model averaged across very different sites can perform worse at each one than a local model would.
- Model updates are not harmless; without secure aggregation and differential privacy, they can be reverse-engineered.
- Agree on ownership of the shared model and on audit rights before training starts.
- Consider simpler options first, since a data-sharing agreement or synthetic data often solves the same problem with less engineering.