Federated learning is a machine learning approach in which a model is trained across multiple decentralised devices or servers holding local data samples, exchanging model updates rather than the raw data itself.
In practiceFederated learning is used where data cannot be centralised for legal, commercial or bandwidth reasons. A coordinator distributes a model, each participant trains on its own data and only parameter updates are aggregated. Practitioners often overstate the privacy guarantees: updates can leak information about training data unless combined with techniques such as secure aggregation or differential privacy. Federated learning is also operationally heavier than classical training because of the orchestration, versioning and validation needed across heterogeneous participants.
A consortium of hospitals trains a clinical decision support model using federated learning so that each hospital keeps patient records on its own infrastructure, with only model weights exchanged via a central aggregator.
This definition is maintained by Moweb partners and used in live client engagements. For how Federated learning applies to your estate, or to challenge a working definition, speak to a partner.