Privacy-preserving collective reinforcement learning using fully homomorphic encryption in usage-based insurance

Usage-based insurance (UBI), a vehicle insurance model where premiums are dynamically adjusted based on real-world usage and driving behaviors, increasingly relies on collective reinforcement learning (CRL) techniques to refine risk assessment and pricing strategies. However, the delegation of computationally intensive vehicle data processing tasks to untrusted cloud servers introduces significant privacy risks, as sensitive information may be exposed to unauthorized entities during computation. To address this critical security challenge, we propose a comprehensive privacy-preserving framework for UBI, beginning with an architectural design that isolates sensitive data from untrusted environments. We then introduce a suite of homomorphic algorithms for reciprocal calculation, scalar product, and sigmoid using the CKKS fully homomorphic encryption (FHE) algorithm. Building upon these foundational operations, we develop homomorphic variants of back propagation and gradient descent algorithms. By leveraging these innovations, we propose a novel privacy-preserving CRL algorithm via FHE that maintains data confidentiality throughout model training and inference, and is then integrated into a UBI system. Simulation results and analysis validate the effectiveness and practicality of our algorithms.