Cross-Vehicle Federated Learning for Unsupervised Scoring of Driver Behavior Under Privacy Constraint

Scoring driver behavior is crucial for profiling drivers to improve driving habits with beneficial feedback, significantly reducing the risks of accidents, fuel consumption, and traffic emissions. Although connected vehicles enable extensive collection of driving data for accurate scoring models, two key challenges persist: 1) the absence of target scores for traditional regression algorithms and 2) privacy concerns that hinder centralized learning (CL) to build unsupervised scoring models. We introduce FedDriveScore, a privacy-preserving unsupervised scoring framework tailored for the Internet of Vehicles context. This solution leverages a mixed distribution approach to form an objective scoring model that utilizes probability distributions and statistical weights to assess driver behavior. Homomorphic encryption-based federated learning (FL) is employed to collaboratively implement the scoring model across multiple vehicles without sharing raw data. To mitigate the detrimental effects of skewed metric distributions, an enhanced FL method is developed to ensure utility consistency between models generated through FL and CL. Experimental results from the two real-world datasets show that the proposed solution achieves remarkable performance, with scoring utility losses of less than 0.0001 (mean-squared error) and 0.01 (mean absolute error), representing approximately a 99% improvement over traditional FL benchmark. Our solution demonstrates high alignment with CL outcomes while effectively profiling drivers in an objective and equitable manner.