Secure Data Engineering Pipelines For Federated Insurance AI: Balancing Privacy, Speed, And Intelligence

Abstract

The insurance underwriting process assesses potential policies, and assigns risk ratings and premiums. In recent years, actuaries have become increasingly interested in leveraging machine learning algorithms to improve existing underwriting models that rely primarily on demographic data such as age, gender, and location for insurance applications such as auto, riders, and renters. Machine learning algorithms can be customized to include a wider array of potentially predictive features and can be more flexible in capturing non-linear relationships between these predictive features and policy loss risk. Moreover, technological advances now also make it feasible to use machine learning algorithms for the underwriting process itself and not only during the rate evaluation.
Traditional machine learning techniques train on pooled historical data. Federated learning is a technique to train machine learning models on separate, private data silos. Training data remains on the devices while information on the model is transferred. In federated learning, this is repeated until sufficient model performance is achieved. Beyond privacy, the training of a machine learning model can also be a matter of speed and intelligence. A federated system that balances data privacy, speed, and intelligence is federated insurance AI. However, federated insurance AI has distinct implementation challenges around performance and security. We aim to address secure data engineering pipelines for practical federated insurance AI. The key is rapid, high-security transformations of data.