Term

Federated Learning

Federated learning is a way to train one shared model across many parties without any of them handing over their raw data. Instead of pooling everything in one place, each participant trains locally on its own data and sends back only model updates (gradients or weights), which a coordinator averages into a global model. Google introduced the idea around 2016 to improve phone keyboards without uploading what people typed, and the same pattern now shows up wherever data is sensitive, regulated, or simply too large to move: hospitals, banks, edge devices.
Reviewed by 7wData

Why it matters

For a governance and data audience, federated learning is interesting because it changes WHERE the risk sits, not whether there is risk. I am seeing it pitched as a privacy silver bullet, and it is not. The raw records stay home, which is real and useful, but the model updates can still leak information about the training data through membership-inference and reconstruction attacks. So federated setups usually need differential privacy or secure aggregation bolted on to actually deliver the privacy promise. It moves the data-governance question from “who holds the data” to “what can be inferred from the gradients”, which is a harder question, not a solved one.

Where you’ll encounter it

Three contexts. A healthcare or financial-services pitch where data residency or regulation forbids centralizing records, and federated learning is how you still train across institutions. A privacy-engineering conversation where the honest follow-up is “federated plus what”, because federated alone is not a privacy guarantee. And an edge or mobile setting where the data is generated on-device and was never economically worth shipping to a central store in the first place.

Part of the 7wData AI Glossary. Tracking how concepts like this move in the expert conversation: daily signals at ins7ghts.com.