AI Trend Research

← AI glossary

AI glossary · Basic · AI security · Data engineering

Federated learning

Federated learning is a method in which data is not gathered in one place; instead, each device or institution trains a model a little on its own data, and only those 'training results (updates)' are collected to build a single shared model.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-09
In plain words It is like several cooks each testing the same recipe at home and, without sending any ingredients, sending only improvement notes such as 'let's use a bit less salt' to headquarters to refine the recipe. However, the fact that the notes alone can hint at the ingredients is a limitation that remains in real federated learning too.

How it works

First, a central server distributes the current shared model to many participants, such as phones or hospitals. Each participant trains the model a little on its own data and sends the server only how the model changed (the update), not the raw data. The server combines these updates into a better shared model and distributes it to the participants again. Repeating this process many times gradually improves the model.

One round of federated learningPhone A → Local training; Phone B → Local training; Hospital C → Local training; Local training → Combine updates(Updates only); Local training → Combine updates; Local training → Combine updates; Combine updates → New shared model; New shared model → Local training(Redistribute)Participant side · data stays putPhone APersonal dataPhone BPersonal dataHospital CMedical dataLocal trainingLocal trainingLocal trainingCombine updatesCentral serverNew shared modelUpdates onlyRedistribute
One round of federated learning Raw data stays with each participant, and only the updates from training are gathered on the server and combined. The new shared model is then distributed back to the participants.

Why it matters

Sensitive data such as personal messages, medical records and financial transactions is hard to gather in one place because of legal or trust issues. Federated learning makes it possible to build a model that learns from large amounts of data while the data never leaves where it originally sits. It opens a way for multiple hospitals or banks to collaborate without handing data to one another. Another advantage is the reduced cost and risk of moving data.

Things to keep in mind

Even when only updates are sent, part of the original data can be inferred from them, so federated learning is usually combined with differential privacy or encryption techniques. Another problem is that participants' data can differ greatly in nature, so training does not progress evenly. Device capabilities and network conditions vary widely, which makes operation complex, and it also has to guard against data poisoning attacks in which malicious participants deliberately send bad updates.

Example

Next-word suggestions on smartphone keyboards are often cited as a typical use of federated learning. Sentences the user types never leave the phone; when the phone has spare capacity, such as while charging, it trains the model a little and sends only the update. In healthcare, it is used by multiple hospitals to jointly train diagnostic support models without sharing patient images.

Related terms

Differential privacy On-device AI Personal data de-identification Data poisoning Data governance