Data stays where it lives. Federated learning trains models across many devices without centralizing raw data. Each device computes updates on local data, sends only those updates to a server, and the server aggregates them into a shared model.

The approach suits situations where privacy or bandwidth makes centralization impractical. Keyboards learn typing patterns without sending keystrokes. Hospitals train diagnostic models without sharing patient records.

Core mechanics

• Local training on each device
• Gradient or model updates sent to a server
• Aggregation into an updated global model
• Distribution back to devices for the next round

Challenges persist. Devices have different data distributions, connection speeds, and availability. Communication overhead is significant, and updates can leak information if not carefully protected. Secure aggregation and differential privacy add safeguards at additional cost.