Federated differentially private framework for risky driving behaviour assessment

Connected and Automated Vehicles (CAVs) offer a transformative opportunity to enhance road safety by identifying and mitigating risky driving behaviours. However, current methods for assessing these behaviours in CAVs overlook privacy concerns, as all terminal devices are required to upload original data directly to a central server for analysis. This data often includes sensitive personal information, raising the risk of potential privacy breaches. To address this challenge, we introduce Federated Learning (FL), a privacy-preserving machine learning technique, and propose a Federated Differentially Private based Risk Assessment Network (FedDPRAN) for risky driving behaviour assessment. In particular, we begin by extracting driving behaviour characteristics for each client using the Risk Assessment Network (RAN), following the adversarial principle of minimizing the difference in driving behaviour features while maximizing the difference in privacy features of the driver. Then, a new FL solution that utilizes the local RAN model is proposed to collaborate and exchange learned parameters with a cloud server without sharing actual data. To enhance privacy in the FL framework, we integrate a Differential Privacy (DP) solution for each client. Comprehensive experiments are conducted based on two real-world datasets. The results demonstrate that our approach to assessing risky driving behaviour in CAVs protects privacy while being superior to state-of-the-art schemes.