Driver Behavior Anomaly Detection Based on Federated Learning Considering Data Distribution Imbalance

This study presents a cross-device federated learning framework for detecting anomalous behavior in automotive mobility and evaluates its performance across various experimental scenarios. The proposed framework retains data locally on vehicle clients, ensuring data privacy while achieving high predictive performance through cross-device federated learning settings. It addresses challenges specific to automotive mobility, such as data distribution imbalance, and employs the lightweight deep learning model like MobileNet for computational efficiency, enabling real-time anomaly detection. Experimental results can demonstrate that the federated learning model achieves accuracy comparable to centralized models without requiring the direct sharing of sensitive driver data. This highlights the framework’s ability to balance data privacy and performance, making it suitable for privacy-sensitive environments such as smart mobility platforms. We believe that the practicality of our framework in mobility applications and its broader potential for developing smart intelligent systems to comply with stringent privacy regulations can offer a valuable solution for integrating artificial intelligence into data-driven industries.