Federated Learning in 2026: How Privacy-Preserving AI Training Is Reshaping Sensitive Industries
How federated learning trains AI without pooling data: the EXAM COVID-19 study across 20 institutions, the 71-site FeTS brain tumor study, Gboard on phones, why it is not private by itself, and when to use it.
Most machine-learning models are trained by collecting data in one place. That is difficult or impossible when data is sensitive, regulated, or commercially valuable, as with hospital records, bank transactions, or what people type on their phones. Federated learning offers a different approach: instead of moving data to the model, it sends the model to the data, trains locally, and combines only the resulting model updates.
This guide explains how federated learning works, what published studies show about its results in healthcare and on mobile devices, why it is not automatically private, and how organizations can decide whether it fits their needs.
How federated learning works
The approach was introduced in a 2016 paper by Google researchers, McMahan et al., “Communication-Efficient Learning of Deep Networks from Decentralized Data”, which proposed the Federated Averaging algorithm. A typical training round works like this:
- A coordinating server sends the current model to participating clients, such as hospitals, banks, or phones.
- Each client trains the model on its own local data for a short time.
- Clients send back model updates (changes to the model’s parameters), not raw data.
- The server aggregates the updates, often by weighted averaging, to produce an improved global model.
- The process repeats for many rounds.
There are two broad settings:
| Setting | Participants | Typical scale | Example |
|---|---|---|---|
| Cross-device | Phones, browsers, or sensors | Thousands to millions of clients, each with little data and intermittent availability | Keyboard prediction |
| Cross-silo | Organizations such as hospitals or banks | A handful to dozens of clients, each with large datasets and reliable infrastructure | Multi-hospital medical imaging |
A comprehensive survey by Kairouz et al., “Advances and Open Problems in Federated Learning”, written by researchers from many institutions, is the standard reference for the field’s techniques and unsolved challenges.
Evidence from healthcare
Healthcare is the most studied cross-silo application because patient data is highly sensitive and often cannot leave institutions under privacy laws and data-governance rules.
EXAM: COVID-19 outcome prediction across 20 institutions
In a study published in Nature Medicine, researchers used data from 20 institutions around the world to train a federated model called EXAM (electronic medical record chest X-ray AI model) that predicts the future oxygen requirements of symptomatic COVID-19 patients from vital signs, laboratory data, and chest X-rays. According to the paper:
- EXAM achieved an average area under the curve (AUC) above 0.92 for predicting outcomes at 24 and 72 hours after initial presentation to the emergency department.
- It improved average AUC by 16% across participating sites and increased generalizability by an average of 38% compared with models trained at a single site on that site’s data.
- The data from the 20 institutions remained under their own custody.
FeTS: brain tumor segmentation across 71 sites
A study in Nature Communications described what its authors called the largest federated learning study to date, involving data from 71 sites across six continents to build an automatic tumor boundary detector for glioblastoma, a rare brain cancer. The dataset included 6,314 cases. The authors reported a 33% improvement in delineating the surgically targetable tumor and a 23% improvement for the complete tumor extent, compared with a model trained on publicly available data.
What these studies show
Both studies suggest that federated learning can produce models that generalize better than single-site models, because they learn from more varied patients, equipment, and clinical practices. They also show that federated learning in healthcare requires substantial coordination: agreeing on data formats, labels, and preprocessing across institutions is often as much work as the training itself.
Evidence from mobile devices
Google described early use of federated learning in its Gboard keyboard on Android: phones store information about suggestions locally, train on that history on-device, and contribute updates to improve the next version of the model. A later paper, Hard et al., “Federated Learning for Mobile Keyboard Prediction”, reported training a next-word prediction model this way.
Cross-device federated learning must handle devices that are often offline, have limited battery and bandwidth, and hold small amounts of data. Training typically happens only when devices are idle, charging, and on unmetered networks.
Federated learning is not private by itself
A common misconception is that because raw data does not leave clients, federated learning is automatically private. Research shows otherwise. Zhu et al., “Deep Leakage from Gradients”, demonstrated that shared gradients can, in some settings, be used to reconstruct training examples. NIST’s blog series on privacy-preserving federated learning also discusses privacy attacks in federated learning.
For that reason, practical systems combine federated learning with additional protections:
| Technique | What it does | Trade-off |
|---|---|---|
| Secure aggregation | Cryptographic protocols let the server see only the combined update, not individual clients’ updates (Bonawitz et al.) | Extra computation and communication; handling dropouts |
| Differential privacy | Adds calibrated noise to updates to limit what can be learned about any individual | Can reduce model accuracy |
| Trusted execution environments | Run aggregation in protected hardware | Depends on hardware trust |
| Access controls and auditing | Limit who can run training and inspect models | Organizational effort |
The privacy-utility trade-off, particularly with differential privacy, is one of the main areas of ongoing research.
Other challenges
- Data heterogeneity. Data across clients is rarely identically distributed. A hospital serving older patients or using different scanners will differ from others, which can slow training or reduce accuracy. Personalized federated learning adapts the global model to local conditions.
- Communication costs. Sending model updates repeatedly can be expensive, especially for large models and many devices.
- Data quality and labeling. Without seeing the data, coordinators must rely on participants to label and preprocess it consistently.
- Poisoning and robustness. Malicious or faulty clients can send harmful updates. Robust aggregation methods and client validation help.
- Governance. Participants need agreements on who owns the resulting model, how it can be used, and how contributions are credited.
Federated learning and data-protection law
Federated learning can make compliance easier, but it does not take a project outside privacy law. Under regimes such as the EU’s GDPR, model updates and trained models can still carry information about individuals, so organizations generally still need a lawful basis for processing, a data-protection impact assessment for high-risk uses, and clear roles for each participant. Treat federated learning as one safeguard in a privacy design, alongside data minimization, security controls, and contracts, rather than as an exemption.
Tools and frameworks
Open-source frameworks make it easier to experiment and deploy:
- NVIDIA FLARE (Federated Learning Application Runtime Environment), widely used in medical imaging research.
- Flower, a framework-agnostic library that works with PyTorch, TensorFlow, and other tools.
- TensorFlow Federated and PySyft, among others, for research and prototyping.
When to consider federated learning
Federated learning is a good fit when:
- Data cannot be centralized for legal, contractual, or competitive reasons.
- Multiple parties would benefit from a shared model trained on more varied data.
- Participants have enough computing capacity and technical staff to run local training.
- There is a clear governance agreement among participants.
It may be unnecessary when data can be lawfully pooled with appropriate safeguards, or when one organization’s data is sufficient. In those cases, centralized training is usually simpler, faster, and easier to debug.
A planning checklist
- Define the shared prediction task and success metric.
- Agree on data standards, labels, and preprocessing across participants.
- Decide on privacy protections: secure aggregation, differential privacy, or both.
- Set governance terms: model ownership, permitted uses, publication, and exit.
- Start with a small pilot among a few participants and compare results with single-site models.
- Monitor for data drift, poisoning, and performance differences between sites.
Related coverage
For related topics, see our articles on AI in financial services, another regulated sector, and enterprise AI governance and compliance.
Bottom line
Federated learning trains models where data lives and shares only model updates. Published studies, including the 20-institution EXAM model for COVID-19 and the 71-site FeTS glioblastoma study, show it can improve generalization over single-site models without pooling patient data. It is not private by itself, because updates can leak information, so real deployments add secure aggregation, differential privacy, and strong governance. For organizations that cannot share data but would benefit from collaboration, it is a practical, though demanding, option.
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
- Dayan et al., “Federated learning for predicting clinical outcomes in patients with COVID-19,” Nature Medicine (2021): https://www.nature.com/articles/s41591-021-01506-3
- Pati et al., “Federated learning enables big data for rare cancer boundary detection,” Nature Communications (2022): https://www.nature.com/articles/s41467-022-33407-5
- Google Research, “Federated Learning: Collaborative Machine Learning without Centralized Training Data”: https://research.google/blog/federated-learning-collaborative-machine-learning-without-centralized-training-data/
- Hard et al., “Federated Learning for Mobile Keyboard Prediction”: https://arxiv.org/abs/1811.03604
- Zhu et al., “Deep Leakage from Gradients”: https://arxiv.org/abs/1906.08935
- Bonawitz et al., “Practical Secure Aggregation for Federated Learning on User-Held Data”: https://arxiv.org/abs/1611.04482
- NIST, “Privacy Attacks in Federated Learning”: https://www.nist.gov/blogs/cybersecurity-insights/privacy-attacks-federated-learning