Normally a model learns from data gathered in one place, which raises privacy concerns. Federated learning flips this: the model is sent to where the data lives, such as many phones, learns locally, and only the resulting improvements, not the raw data, are combined centrally.
This lets a keyboard or health app get better from real use without collecting sensitive personal information in a single store. It is more complex to coordinate and the shared updates must themselves be handled carefully to avoid leaking hints about the underlying data, but it offers a way to improve models while keeping personal data on the owner's device.