federated learning (FL)

advancedspreading

A training method in which the AI model learns locally on many devices or at many institutions, and only what it has learned is shared centrally, not the raw data.

In plain terms

In traditional machine learning, data is collected in one place and the model is trained on it there. Federated learning reverses this: the data does not go to the model, but the model goes to the data. Your phone, hospital, or bank trains a copy of the model locally and only sends back what it changed.

Imagine a hundred chefs trying to perfect a recipe together, but none reveal who they are cooking for or what the guests eat. Each tries the recipe at home, then submits only: 'less salt, bake five minutes longer.' Someone averages the suggestions, and everyone receives the improved recipe. The guest book stays in the kitchen the whole time.

This is useful where data is sensitive or legally cannot be merged: medical records, bank transactions, messages typed on your phone. This way, even actors who would never share their data with each other can build a joint model.

How it works

A central server sends the current version of the model to participants, meaning phones, hospitals, and companies.
Each participant continues training the model for a short time on its own locally stored data.
Participants do not send back the data but the model modifications, meaning how and by how much the internal numerical values (weights) shifted.
The server aggregates, typically averages, the received modifications, and from this the model's new, shared version is created.
The cycle often repeats until the model is good enough; modifications are often protected with encryption or by adding intentional noise.

An everyday example

Your phone's keyboard is increasingly accurate at predicting your next word. At several manufacturers, such as Google's Gboard, this is achieved by the device learning locally from your typing when it is not in use, plugged in, and connected to Wi-Fi (typically at night). It sends only model improvements, not your messages.

Who is affected

patient, doctorMultiple hospitals can train a shared diagnostic model without patient data ever leaving any single institution. → Health

bank customer, investorBanks can jointly develop a fraud detection model without sharing their customers' transactions with each other. → Money

lawyer, citizenIt can help with data protection compliance, but does not exempt you from the GDPR on its own, as model modifications can also carry personal data. → Law

developerIt requires a completely different architecture than central training: you must work with unreliable connections, different devices, and data that is not visible. → Code

A common misconception

MythIf the model is prepared with federated learning, my data remains fully secure, with no risk of leakage.

In realityThe raw data does indeed remain in place, but certain attacks can infer part of the original data from the returned modifications. Therefore, it is usually used together with supplementary protection, such as differential privacy and secure aggregation.

Limits and risks

Participants' data can vary greatly, which may cause the shared model to learn more slowly and be less accurate than one trained centrally.
A malicious participant could intentionally send distorted modifications, thereby ruining or manipulating the shared model.
It requires much communication and local computing capacity, which becomes a bottleneck on weak devices or slow networks.
Since no one sees the entire dataset, it is harder to verify data quality and filter out biases.

Similar, but not the same

distributed training – This is the broader concept: training occurs across multiple machines. In its traditional form, the data belongs to a single owner who freely distributes it across their own servers, whereas in federated learning, the data stays with its owner.
differential privacy – This is a mathematical method that intentionally adds noise to limit how much can be revealed about an individual's data; it reduces risk but does not eliminate it. It complements federated learning rather than replacing it.
on-device AI – This simply means the model runs locally on your device, whereas federated learning describes how the model learns from data across many devices.

Related concepts

Broader

machine learning, privacy-enhancing technologies, distributed training

Narrower

federated averaging, cross-device federated learning, inter-institutional federated learning

On the edge

secure aggregation, differential privacy, on-device AI, edge computing, synthetic data, training data

A brief history

The term was introduced by Google researchers, Brendan McMahan and colleagues, in a study published in 2016, which also described the federated averaging method. Google publicly presented it in 2017, using the Gboard keyboard as an example.

Where things stand

On mobile phones it already runs widely in the background, for example in keyboards and speech recognition. In healthcare and the financial sector, live deployments are increasingly common, but many projects are still experimental.

Status: 2026 November

Also known as

federated learning · federated learning · federated learning · consolidated learning · federated learning · FL · collaborative learning · federated machine learning

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.