A Cognitive-Intelligence-Based Personalized Federated Approach for Monitoring Driver Behavior

To enhance user safety and experience, next-generation transportation system needs to understand and classify drivers’ behavior accurately. For instance, reckless or unsafe behavior may lead to road accidents. Traditional driver behavior monitoring methods require local data to be uploaded to a central server for model training, which is impractical due to the large data size and privacy concerns. This paper presents Federated Learning (FL) as a method for enabling local training on clients while keeping their data private and minimizing communication overhead. However, the diversity in drivers’ skills, habits, and unique preferences adds significant complexity to modeling driver behavior. To address this, we introduce pFedCI, a novel personalized FL framework based on cognitive intelligence for monitoring driver behavior. In this framework, each client retains a locally trained centralized model, which fine-tunes the global model to create distinct and personalized models. A key innovation of the pFedCI framework is its ability to allow clients to choose the optimization algorithm during each training round adaptively. Additionally, the server optimizes the global model to ensure its generalizability. Extensive experiments on the UAH and KIA Soul datasets show that pFedCI significantly outperforms previous FL methods. Moreover, we demonstrate that pFedCI achieves better generalization across various state-of-the-art models and clients with diverse features.