Federated learning is a machine learning technique in which many devices or organizations train a shared model together while keeping their data where it is. Instead of collecting everyone’s data in one place, a central server sends the current model to participants, each participant improves it using local data and only the resulting model updates travel back. The server combines the updates into a better shared model. The approach lets companies and institutions learn from sensitive data, such as phone typing or hospital records, with less exposure than traditional centralized training.
History of federated learning
The term was introduced in 2016 by researchers at Google, who published a paper by H. Brendan McMahan and colleagues describing communication-efficient learning from decentralized data. They wanted to improve predictive keyboards on Android phones without uploading what people typed. Their federated averaging algorithm, known as FedAvg, became the foundation of the field. Since then, researchers have extended the idea to healthcare, finance and the internet of things, and open-source frameworks such as TensorFlow Federated, Flower and PySyft have made it easier to experiment. A major survey in 2019, Advances and Open Problems in Federated Learning, set out the research agenda.
How federated learning works
Training proceeds in rounds.
- The server starts with an initial global model.
- It selects a group of clients, such as phones that are charging and on Wi-Fi, and sends them the model.
- Each client trains the model for a few steps on its own data.
- Clients send back their updated weights or gradients, not the data.
- The server averages the updates, often weighting by how much data each client has, to form a new global model.
- The process repeats until the model is good enough.
Because communication is the main bottleneck, FedAvg lets clients perform several local training steps before sending an update, which cuts the number of rounds needed.
Types of federated learning
- Cross-device: a very large number of unreliable consumer devices, such as phones, each with a small amount of data. Clients may drop out at any time.
- Cross-silo: a small number of organizations, such as hospitals or banks, each with large datasets and reliable computing. Participants are known and can be contractually bound.
- Horizontal: participants have the same features but different individuals, such as two hospitals with the same types of records for different patients.
- Vertical: participants hold different features about the same individuals, such as a bank and a retailer that serve the same customers.
Privacy and security
Keeping data local is a major privacy gain, but model updates can still leak information, since researchers have shown that attacks may reconstruct training examples from gradients. Federated learning is therefore combined with other protections. Differential privacy adds carefully calibrated noise so that no single person’s data changes the result much. Secure aggregation uses cryptography so that the server sees only the sum of updates, not those from any one device. Trusted execution environments and homomorphic encryption add further safeguards. Malicious participants can also try to poison the model by sending bad updates, so robust aggregation and anomaly detection are used to resist them.
Applications of federated learning
The first large deployment was next-word prediction and emoji suggestion on mobile keyboards. Phone makers use it to improve voice assistants and wake word detection without uploading recordings. In healthcare, hospitals have trained models on medical images and records across institutions, which helps rare disease research where no single hospital has enough cases. Banks use it for fraud detection across institutions without sharing customer data, and in manufacturing and energy it helps combine sensor data from many sites. Autonomous vehicle makers explore it to learn from fleets of cars. Regulations such as the European General Data Protection Regulation make approaches that limit data movement attractive.
Advantages of federated learning
The key benefits are better privacy, since raw data stay on the device, and compliance with rules that restrict data transfer. It reduces the need to move and store huge datasets, and it lets models learn from data that would otherwise be unavailable. Models can also be personalized to each user while still benefiting from the whole group. For cross-silo uses, it allows competitors or institutions to cooperate on a shared model without exposing business secrets.
Challenges of federated learning
Several problems remain. Data across clients are typically not identically distributed, since one person’s typing differs from another’s, which can slow training and hurt accuracy. Devices differ in speed, memory and connectivity, and many drop out mid-round. Communication is costly, so compression and fewer rounds are important. It is harder to debug and evaluate a model when the developer cannot inspect the data. Fairness is a concern, because devices that participate more often may dominate the model. Finally, the privacy guarantees depend on correct use of extra safeguards, and a federated system is not automatically private.
Federated vs. centralized and distributed learning
In centralized learning, all data are gathered in one location. In distributed learning inside a data center, data are split across machines for speed, but the data owner controls everything and the splits are typically uniform. Federated learning differs because the data are owned by different parties, are not under central control and are not evenly distributed, and privacy is a central goal rather than a side effect.
A worked example: a predictive keyboard
Consider the next-word suggestions on a phone keyboard. A global model is stored on a server. When a phone is idle and charging, it downloads the model and trains briefly on what the user has recently typed, which never leaves the device. The phone sends a small encrypted update, and the server combines updates from thousands of phones. The new global model is then pushed back to devices. Over time the suggestions improve for everyone, including for new slang that no developer typed into a dataset, and the company never stores a person’s messages. The example shows why federated learning suits personal data and why engineers also use noise and encryption to protect the updates.
The future of federated learning
Research is moving toward personalization, where each client keeps a model tuned to its own data, and toward federated fine-tuning of large pretrained models, which lets organizations adapt a foundation model without sharing their documents. Work on fairness, efficiency and stronger privacy proofs continues, and standards bodies and regulators are examining how federated methods fit into data protection law. As devices become more powerful and privacy expectations grow, the approach is likely to spread across health, finance and consumer technology.
FAQs
Does federated learning protect privacy completely? No. It reduces exposure of raw data, but updates can leak information, so it is combined with differential privacy and secure aggregation.
Who invented federated learning? Researchers at Google, led by H. Brendan McMahan, introduced the term and the FedAvg method in 2016 and 2017.
Is federated learning used on my phone? Possibly. Some keyboard and voice features on phones are improved using federated methods.
More in Machine Learning
Key facts
- Google researchers introduced the term federated learning and the FedAvg algorithm in 2016 and 2017.1
- In each round, devices train on local data and send only model updates, which a server averages into a new global model.1
- Federated learning is often combined with differential privacy and secure aggregation to strengthen privacy.2
Related concepts
Quick checkWhat is shared in federated learning instead of raw data?Show answer
Model updates, such as changed weights or gradients, computed locally on each device or site and then aggregated by a server.
Sources
- McMahan et al., Communication-Efficient Learning of Deep Networks from Decentralized Data, https://arxiv.org/abs/1602.05629
- Kairouz et al., Advances and Open Problems in Federated Learning, https://arxiv.org/abs/1912.04977
- Google AI Blog, Federated Learning: Collaborative Machine Learning without Centralized Training Data, https://research.google/blog/federated-learning-collaborative-machine-learning-without-centralized-training-data/
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
