Federated learning
Also called collaborative learning, decentralised machine learning
Federated learning is a way to train one machine learning model across data held by several organisations without moving or pooling that data. Each party trains the model locally and shares only model updates. Banks use it to detect fraud and money laundering patterns that span institutions while keeping customer data in-house.
How it works
A coordinator sends a starting model to each participating bank. Each bank trains it on its own transaction data and returns only the changed model parameters. The coordinator combines these updates into an improved shared model and sends it back. The cycle repeats. Techniques such as secure aggregation and differential privacy limit what any party can learn about another's customers from the updates.
Criminals spread activity across banks so no single institution sees the full pattern. Data protection and banking secrecy rules restrict sharing raw customer data. Federated learning offers a route between the two. The UK data protection regulator lists it as a privacy-enhancing technology, while noting that model outputs can still reveal personal data and other risks remain.
The rules
Requires measures such as data minimisation to be built into processing, which privacy-enhancing technologies can support.
Controllers processing personal data in the EUDescribes federated learning as a privacy-enhancing technology and explains its data protection benefits and residual risks.
Organisations processing personal data under UK GDPRCommon questions
Does federated learning share customer data?
No raw data leaves each organisation. Only model updates are shared. Those updates can still leak information in some cases, so firms add protections such as secure aggregation and differential privacy.
Why would banks use federated learning for AML?
Money laundering networks use accounts at many banks. A model trained across several banks' data can spot cross-bank patterns that no single bank's model would see, without pooling customer records.
Is federated learning anonymisation?
Not on its own. The ICO treats it as a technique that reduces the personal data shared. Model outputs can still reveal personal information, so data protection law continues to apply.
