Federated Learning
The most useful training data — your messages, photos, health logs — is exactly the data that should never leave your device. Federated learning resolves the conflict by inverting the flow: instead of shipping data to the model, the model ships to the data. Each device trains a local copy on its own data, sends back only the resulting weight updates, and a server averages the updates from thousands of devices into an improved shared model that travels back down. Raw data never moves; updates can still leak hints, so they are commonly noised (differential privacy) or aggregated so no single device's contribution is readable.
Watch the four-phase loop: local training, updates travelling up, averaging, the improved model travelling down — the green data dots never move. Switch to centralized to see the contrast: raw data leaves every phone for one honeypot server.
Federated learning inverts it: the model is sent to the data. Each phone trains on its own data locally, sends back only weight updates, and the server averages them into a better shared model. This is how phone keyboards learn new slang without reading your messages.
Check yourself
The raw data never leaves the phone, yet federated learning alone is not perfectly private. What is still being shared, and why can it leak?
Go deeper (free): Google AI — Federated learning comic ↗