what is federated learning
Federated learning is a way of training AI models across many separate devices or locations without ever centralizing the underlying data — each device trains on its own local data and shares only the resulting model updates, which get combined into a single improved model while the raw data itself never leaves its original location.
```flowchart
{
"rows": [
{ "type": "nodes", "items": [
{ "title": "Data stays on local devices", "sub": "e.g. phones, hospitals, individual organizations", "tone": "blue" }
]},
{ "type": "arrow" },
{ "type": "group", "title": "Each device trains locally, shares only updates", "items": [
{ "title": "Model updates sent, raw data never leaves", "sub": "preserves data privacy and locality", "tone": "green" }
]},
{ "type": "arrow" },
{ "type": "nodes", "items": [
{ "title": "Updates combined into one improved model", "sub": "central model benefits from all data without centralizing it", "tone": "orange" }
]}
]
}
```
**Federated learning exists to let an AI model benefit from data spread across many separate locations without requiring that data to ever be pooled together in one place.** Traditional AI training typically requires centralizing all training data in one location, which can be impractical or undesirable when the data is sensitive, regulated, or simply too distributed to move — federated learning instead sends the model to where the data already lives, trains locally, and brings back only the resulting model updates, letting the central model improve without the underlying data ever being centralized or exposed.
```svg
```
```svg
```
| Aspect | Centralized training | Federated learning |
|---|---|---|
| Where training happens | Central server with pooled data | Locally, on each device or site |
| What travels | Raw training data | Model updates only |
| Data privacy | Data leaves its original location | Data never leaves its original location |
| Common use cases | General-purpose model training | Mobile devices, healthcare, regulated industries |
**Federated learning has become particularly important in domains where data privacy or regulation makes centralizing data difficult or legally restricted.** Healthcare, finance, and other regulated industries often hold data that can't easily be moved or pooled due to privacy laws or organizational policy, making federated learning an attractive way to still benefit from training across that data without requiring it to be centralized in a way that regulation or policy would forbid.
**Coordinating federated learning across many devices introduces distinct engineering challenges compared to traditional centralized training, particularly around communication and consistency.** Because updates come from many independent devices that may have inconsistent connectivity, different amounts of local data, or different hardware capabilities, federated learning systems need specific techniques to combine those updates fairly and efficiently — this coordination challenge is a major reason federated learning remains an active area of both research and engineering.
**Federated learning doesn't eliminate every privacy concern by itself, since model updates can sometimes still leak information about the underlying local data.** Even though raw data never leaves the device, researchers have shown that model updates can sometimes be analyzed to infer information about the data that produced them — this is why federated learning is often combined with additional privacy techniques, rather than treated as a complete privacy solution on its own.
Read federated learning through a train-locally-share-only-the-lesson lens: rather than gathering everyone's raw data in one place, it lets each device or location learn from its own data and share only what it learned — letting a shared model improve from a wide range of data without that data ever having to leave home.