Privacy enhanced data aggregation based on federated learning in Internet of Vehicles (IoV)

As the Internet of Vehicles (IoV) environment continues to evolve, artificial intelligence (AI) technologies have been utilized to provide several applications such as traffic flow prediction and vehicular object detection. Cyberattacks are on the rise since aggregated datasets from multiple vehicles may contain a significant amount of information related to privacy. Federated learning (FL) has gained more attention as it enables data training on local devices without sharing the actual datasets. However, there are still challenges to security, privacy, and heavy communication overhead when applying FL to the IoV. To cope with these issues, we propose a new privacy-enhanced data aggregation scheme based on FL. We apply not only additive secret sharing (ASS) with homomorphic encryption (HE) but also asymmetric encryption such as RSA to guarantee the privacy and security of aggregation results. In our proposed scheme, a roadside unit (RSU), which serves as a group leader, groups vehicles and then chooses qualifying vehicles to not only ensure efficiency with high accuracy but also to reduce communication overhead and improve the scalability of a large number of vehicles. After the initial training that involves all RSUs and selected vehicles, only a randomly selected group is able to update the global model parameter by reflecting on training their local model. It is also helpful not only to reduce communication overhead while keeping acceptable prediction errors but also to prevent continuous receiving from malicious vehicles. The results of security and performance evaluations demonstrate that our proposed scheme is effective in enhancing privacy and efficient in operation.