For Heads-of · Practitioner
Federated Learning
A machine learning approach in which a model is trained across many separate data sources without the data itself being gathered in one place.
- federated learning
- distributed
- privacy
- training
Process
Instead of sending data to a central server, the model is sent to the data holders, trained locally, and only the model updates are returned and combined. The central party never sees the raw data.
Privacy benefit
Sensitive data stays where it is, which reduces privacy risk and can make it possible to learn from regulated data, such as health records or financial information, that cannot lawfully or practically be pooled. It does not remove data protection obligations, and model updates that cross borders still need assessing; see cross-border AI data transfer risk.
Challenges
It is technically complex and expensive to run. Model updates can still leak information about the training data, so strong privacy needs differential privacy on top. And because the central party cannot inspect the data, a malicious participant can poison the shared model, the data and model poisoning risk.
Related Terms
Data Poisoning
An attack in which an adversary deliberately corrupts data an AI model learns from, to degrade it or plant hidden behaviour. It now reaches beyond training data to fine-tuning sets, retrieval corpora and agent memory, and a poisoned model can pass ordinary evaluation.
Differential Privacy
A mathematical framework for limiting what can be learned about any individual from an algorithm's output, by adding carefully calibrated noise so that the output barely changes whether or not that individual's data was included.
Training Data
The dataset used to train a machine learning model, from which the model learns patterns and relationships.
Related risks and controls
Cross-border AI data transfer risk
Keeping data local is the point of the technique, and the transfer question does not disappear because gradients rather than records cross the border.
Data and model poisoning
Every participant is a potential poisoning source, and the aggregation step is where it enters the shared model.
Data minimisation and lawful basis review
The technique reduces what is shared; it does not settle the lawful basis for training on it.