Federated Learning
Last updated June 14, 2026
What is Federated Learning in simple terms?
In simple terms, federated learning teaches a model without gathering everyone's data in one place. Each phone or hospital learns from its own data and shares only the lesson, not the data — keeping the private stuff private.
What is Federated Learning?
Federated learning is a way of training a shared machine learning model across many devices or organizations without collecting their data in one place — each keeps its data locally and sends back only what it learned, so the raw data never leaves home.
Normally, training a machine learning model means hauling all the data into one central place and learning from the pile. Federated learning flips that. The data stays where it was created — on your phone, in a hospital's servers, on a bank's systems — and the *model* travels to the data instead. A copy of the model is sent out to each device, it trains a little on the local data, and then sends back only its updates: a summary of what it learned, never the underlying data itself. A central server gathers these updates from many devices, blends them into an improved shared model, and sends the better version back out. Repeat that cycle many times and you get a model trained on everyone's data without anyone's data ever leaving their device.
The driving reason is privacy, and the contexts where it shines make that obvious. Your phone's keyboard can improve its next-word suggestions by learning from how you actually type, without your messages being uploaded anywhere. Hospitals, which legally and ethically can't pool raw patient records, can still collaborate on a shared diagnostic model by each training locally and sharing only model updates. The neat reframing is this: instead of bringing the data to the model, you bring the model to the data, and only lessons travel back. Think of a group of specialists who can compare notes on what they've learned but are forbidden from sharing their patients' actual files — federated learning is the technical version of that arrangement.
It's not a privacy silver bullet, and that's worth being honest about. The model updates that get shared can, in some cases, leak hints about the underlying data, so federated learning is usually paired with extra protections — techniques that add mathematical "noise" to the updates, or that combine many devices' updates in a way no single one can be read out. It also brings real engineering headaches: phones drop offline, vary wildly in power, and hold very different data, all of which makes coordinating the training harder than centralizing it. So federated learning is best understood as a strong, increasingly important *tool* for training on sensitive, distributed data — not a guarantee of privacy on its own, but a meaningful step toward learning from data you're not allowed to collect.
Real-world example of Federated Learning
Picture a network of independent bakeries that all want a better bread-spoilage predictor, so each can pull loaves before they go stale. The data that would train it — every bakery's exact sales, recipes, and waste logs — is commercially sensitive, and no bakery will hand its books to a rival or a central office. With federated learning they don't have to. A shared prediction model is sent to each bakery's own computer, where it trains quietly on that bakery's private records overnight. In the morning, each bakery sends back only the *adjustments* it learned — not a single sales figure — to a coordinator that merges them into a smarter shared model and redistributes it. Over many rounds, every bakery ends up with a predictor sharpened by the whole network's experience, while each one's books stay locked in its own back office.
Related terms
Frequently asked questions about Federated Learning
What is the difference between federated learning and traditional centralized learning?
In traditional centralized learning, all the data is gathered into one place — a central server or data center — and the model is trained on the combined pile. In federated learning, the data never moves: it stays on each device or with each organization, the model is trained locally, and only the learned updates are sent back to be merged. Centralized is simpler and often faster but requires collecting everyone's raw data; federated keeps the raw data private and distributed, at the cost of more complex coordination. The dividing question is whether the data can leave home. **2. Mechanism — How does federated learning work?**
How does federated learning work?
A central server sends a copy of the current model to many participating devices. Each device trains the model briefly on its own local data and sends back only the resulting updates — the changes to the model, not the data. The server combines all these updates into an improved model and sends the new version back out. This round repeats many times, the shared model getting better each cycle, having effectively learned from everyone's data without that data ever being collected centrally. Extra privacy techniques are often added so the updates themselves reveal as little as possible. **3. Application — What is federated learning used for?**
What is federated learning used for?
It's used wherever the data is valuable for training but too private or too sensitive to centralize. The best-known case is mobile keyboards and phones improving features like next-word prediction from real usage without uploading your messages. Healthcare uses it to let hospitals jointly train diagnostic models without sharing patient records, and banks use it to build fraud models across institutions without exposing customer data. In general, it fits any setting where regulation, competition, or trust forbids pooling the raw data but everyone would still benefit from a model trained on all of it.