What does Federated Learning mean?

Federated learning is a training method in which several organizations jointly improve a model without handing over their own raw data. Training happens locally within each organization, and only the changed model parameters are exchanged. Unlike central training, no shared collection of the underlying data ever forms at any point in the process, neither on the server nor among the participants.

A server first distributes a starting model to all participants. Each organization trains it for a number of rounds on its own data and sends back only the changed weights. The server averages these contributions, weighted by data volume, and redistributes the updated model. Because weights can reveal something about the underlying training data, methods such as differential privacy or encrypted aggregation are added to the process.

Federated learning pays off wherever several organizations share the same question but are not legally allowed to pool their data. Examples include medical analyses across several clinics, fraud detection across several banks, and model improvement on end-user devices. If all the data already sits with a single party, central training remains simpler and cheaper.

The advantage over a shared data pool is that each organization retains sovereignty over its own data. That removes the most demanding part of the coordination effort: the legal clearance of a data transfer. The joint model still reaches a data basis that none of the participants could build up alone.

What the legal department handles for a shared data pool, coordinated operations handle for federated learning. All participants need comparable data structures, well-maintained local technology, and a bindingly agreed training schedule. And because the server merges the contributions, it must be clear beforehand who operates it.