A A Personalized Federated Learning Framework for Post-Event Forensic Traffic Analysis in Autonomous Vehicle Systems

A Personalized Federated Learning Framework for Post-Event Forensic Traffic Analysis in Autonomous Vehicle Systems

Abstract

With the growing prevalence of autonomous vehicles (AVs) in modern transportation systems, exploring post-incident forensic analysis into their operational data is becoming increasingly important for liability evaluations and traffic safety studies. But tough privacy laws, exclusive control over data ownership and proprietary platform architectures all make it challenging for different AV entities to gain hands-on access to raw sensor and telemetry data. To tackle these issues, in this paper we propose a privacy-preserving federated learning framework designed for the post-event forensic traffic analysis in an autonomous vehicle system. The potential of the proposed method lies in that manufacturers, infrastructure providers and regulatory agencies can collaborate an intelligence attack without exhibiting or exchanging any type of sensitive local data to preserve the data privacy and regulation rules. The network is a spatiotemporal deep learning model, which incorporates temporal, spatial and attention mechanism to effectively restore vehicle trajectories as well as identify abnormal driving behavior in intricate traffic scenes. In addition, we propose a client-specific adaptation strategy to adapt to the diversity of AV platforms and traffic patterns for personalized learning while not compromising global model performance. In order to facilitate scalability and deployment opportunity, we employ model compression scheme for minimizing communication overhead during federated updates. Experimental results performed on simulated and real AV datasets show that the proposed approach can simultaneously achieve robust trajectory reconstruction, effective anomaly detection with strong privacy guarantee and communication efficiency. Quantitative results also determine an improvement of around 15% in trajectory prediction accuracy over standard FedAvg, alongside nearly 30% reduction in communication overhead.