Federated Learning

What is federated learning?

Normal training gathers all the data into one place, which is hard for sensitive or scattered data: hospitals can't hand over patient records, and phone makers don't want to upload your data. Federated learning inverts this — data stays local, and only the model's learned updates travel back to a central server to be merged. Everyone trains one model together without anyone surrendering raw data.

How does it work?

Send the model out
The server first ships an initial model to each participant.
Train locally
Each party trains that model a little further on its own local data.
Share only updates
Each sends back just the change in model weights — not a single piece of raw data.
Aggregate the updates
The server combines the updates (say, a weighted average) into a better global model and ships it back out. Rinse and repeat.

What problems does it solve?

Privacy
Raw data never leaves the device, so sensitive information is far less likely to leak.
Data silos
Data scattered across hospitals, banks and phones can be trained on "indirectly" together, without pooling it first.
Compliance
With rules like GDPR tightening, federated learning offers a path that stays legal.

What are its challenges?

Devices are flaky in compute and connectivity, and each party's data isn't identically distributed, so training converges slower and is trickier. Communication overhead and poisoning attacks from bad actors also need their own defenses.

Bottom line: federated learning keeps data at home while the model learns everywhere — training together without giving up anyone's privacy.