Also known as: federated training · on-device training
Federated learning is a technique for training a shared model across many devices without centralising the training data: each device computes updates locally, and only those model updates — not the underlying data — are sent to a server for aggregation.
It allows models to learn from data that cannot legally or practically be centralised, such as text typed on phones or patient records held by separate hospitals.
Model updates are not automatically private. Research has shown that training data can sometimes be partially reconstructed from gradients, which is why serious deployments combine federated learning with secure aggregation and differential privacy rather than relying on decentralisation alone.
Differential Privacy
Differential privacy is a mathematical framework that adds carefully calibrated statistical noise to data or query results, providing a provable guarantee that the presence or absence of any single individual has negligible effect on the output.
Edge AI
Edge AI is the practice of running AI models directly on end-user devices such as phones and laptops rather than on remote servers, keeping input data local to the device.
Data Minimization
Data minimisation is the principle that an organisation should collect and retain only the personal data actually necessary for a stated purpose, and no more, for no longer than needed.
On-Device Processing
On-device processing means performing computation locally on a user's own hardware rather than transmitting data to a remote server, so that the raw input never leaves the device.