One-Sentence Definition
Federated learning is a machine learning technique that trains a model across multiple devices or institutions by sharing only model updates -- not raw data -- so that sensitive information never leaves its original location.
How It Works
In traditional machine learning, training data is collected into a central server, where the model is trained in one place. Federated learning flips this pattern. The model goes to the data, not the other way around.
The process works in rounds. A central server sends the current model to participating devices or nodes. Each node trains the model on its local data and computes an update -- the gradients or weight changes that would improve the model. Only these updates are sent back to the server, which aggregates them (typically by averaging) to produce an improved global model. The raw data on each device is never uploaded.
Google pioneered federated learning at scale in 2017 with its Gboard keyboard. Millions of Android phones train the next-word prediction model locally on each user's typing patterns. The aggregated updates improve predictions for everyone, but Google never sees the individual text anyone types. Apple uses a similar approach for its on-device suggestions in iOS, including QuickType keyboard predictions and Siri improvements.
The basic approach has important refinements. Differential privacy adds calibrated noise to the updates so that the server cannot reverse-engineer any individual's data. Secure aggregation uses cryptographic techniques so the server can only see the combined update, not any single device's contribution. Federated analytics extends the concept beyond model training to general data analysis -- computing aggregate statistics without collecting individual records.
Challenges remain. Devices have different amounts of data (data heterogeneity), unreliable network connections, and varying compute capabilities. Communication costs can be high when models are large. And while federated learning reduces privacy risk, sophisticated attacks can sometimes extract information from model updates, which is why differential privacy is increasingly treated as a requirement rather than an option.
Why It Matters
As privacy regulations like GDPR, HIPAA, and the EU AI Act tighten, organizations need ways to train AI without centralizing sensitive data. Federated learning is a direct answer. Hospitals in a consortium can collaboratively train a diagnostic model without sharing patient records -- NVIDIA's Clara platform enables exactly this. Financial institutions can build fraud detection models that learn from transaction patterns across banks without exposing customer data.
In 2026, federated learning is expanding beyond mobile keyboards and healthcare. Intel and other chipmakers are building federated learning into edge computing hardware. Companies like Owkin apply it to drug discovery, training models on clinical trial data distributed across pharmaceutical partners.
Key Takeaway
Federated learning trains AI models by bringing the computation to the data instead of the data to the computation, enabling collaborative model improvement without compromising privacy.
Part of the AI Weekly Glossary.