What is federated learning?
Federated learning trains a shared model using contributions from multiple devices or institutions without centralizing all raw data.
How it works and where it is used
Participants compute locally and an aggregator combines model updates. Potential uses include mobile keyboards and cross-institution research. Updates can still leak information, data distributions differ, and devices may be offline. Keeping raw data local does not automatically guarantee privacy.
Limits, safety and selection criteria
Secure aggregation and sometimes differential privacy can add protection, with trade-offs in accuracy and communication. Consent, purpose limits and detecting malicious updates are governance issues. Evaluate performance under real deployment conditions, not just a lab metric. Several hospitals may train local updates without sending patient records to a central repository. A coordinator aggregates those updates into a shared model. This does not eliminate the possibility of information leaking through updates; secure aggregation, access controls and sometimes differential privacy need separate evaluation. Different patient populations may produce strong results at one hospital and weak ones at another. Participants also differ in bandwidth and compute capacity. Measure fairness across institutions and data governance alongside overall accuracy.