Federated learning

Federated learning enables the training of a collective model from data distributed across multiple nodes, devices or organizations, without centralizing raw data. Each participant trains a local copy of the model on its own data, then transmits only the weight updates, called gradients, to an aggregation server. This server combines the received gradients to improve the global model, which is then redistributed. Raw data never leaves its originating environment. In actuarial science and insurance, federated learning allows several companies to form a mutual scoring model without sharing their confidential portfolios, or a reinsurer to build a joint catastrophe model with its cedants. Limitations are well documented: gradients can themselves, under certain conditions, be inverted to partially reconstruct training data, which justifies combining federated learning with differential privacy. Moreover, gradient poisoning attacks constitute a vulnerability vector specific to this architecture. European financial regulation, notably DORA, encourages actors to favor this type of architecture for sensitive models.

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, transmits encrypted gradients to a neutral aggregator, and receives back a consolidated model that outperforms any individual model. No personal data leaves the perimeter of each company.

federated learning, apprentissage fédéré, entraînement distribué préservant la vie privée, federated ML