Federated learning
A training method that improves a model on distributed data without the data ever leaving its source, which keeps that data confidential.
McMahan and coauthors at Google described the FedAvg algorithm in February 2016 (published at AISTATS in 2017), and from 2017 the company applied it to Gboard's next-word prediction: keystrokes stay on the phone and only model updates travel. Federated learning trains a collective model on data spread across multiple nodes, devices or organizations, without centralizing raw data. Each participant trains a local copy of the model on its own data, then sends only its weight update, computed after several local training steps, to an aggregation server. The server averages the updates it receives to improve the global model, then redistributes it. Raw data never leaves its originating environment. In actuarial science and insurance, several companies can thus form a mutual scoring model without sharing their confidential portfolios, and a reinsurer can build a joint catastrophe model with its cedants. Limitations are well documented. Under certain conditions, gradients can be inverted to partially reconstruct training data, which is why federated learning is combined with differential privacy. The architecture also has its own weak point: gradient poisoning.
Textbook case: five European insurers wish to improve their auto fraud detection model without sharing customer data. They implement a federated learning scheme: each insurer trains the model locally on its claims, sends its encrypted updates to a neutral aggregator, and receives back a consolidated model that outperforms any individual model. No personal data leaves the perimeter of each company.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.