Definition
Federated learning trains a shared model from computation performed where the data resides. A coordinator distributes a model to selected participants. Each participant trains locally, sends an update rather than raw examples, and the coordinator aggregates those updates into the next shared model.
Federated averaging, or FedAvg, combines several local optimization steps with weighted model averaging. Real deployments must handle participants that are intermittently available, slow, malicious, or trained on data that differs sharply from other participants' data. Communication can be a larger constraint than local computation.
Name and attribution
H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas explicitly wrote "We term this decentralized approach Federated Learning" in their 2016 paper. They presented iterative model averaging for mobile-device data and measured communication-round reductions against synchronized stochastic gradient descent.
Distinguish it from nearby terms
Federated learning reduces raw-data centralization. It is not the same as differential privacy, secure aggregation, or encryption, although those controls can be added. Model updates can leak information, and the coordinator still needs defenses against poisoned or malformed contributions.
Check your understanding
A phone keeps its text locally but sends an unprotected gradient update to the coordinator. Which privacy benefit has federated learning provided, what can the update still reveal, and which additional controls would you consider?