Federated World Models for Privacy-Preserving Collaborative Autonomous Driving in Edge-Vehicle Networks

Abstract

The development of safe and robust autonomous driving systems depends critically on the ability to perceive, predict, and plan over complex and dynamic environments. Traditional centralized approaches require the aggregation of vast amounts of sensitive trajectory and visual data from vehicles, raising significant privacy, security, and regulatory concerns. This paper introduces the concept of federated world models, a distributed learning framework that combines the representational power of latent dynamics models with the privacy-preserving properties of federated learning, operating within edge-vehicle networks. Unlike conventional federated learning that focuses on supervised tasks, federated world models enable vehicles to collaboratively learn a shared generative model of the environment—including scene understanding, motion prediction, and counterfactual reasoning—without exposing raw sensor data. We present a comprehensive system architecture where vehicles serve as local clients, edge nodes perform hierarchical aggregation and differential privacy budgeting, and a cloud server maintains a global world model. The paper examines structural trade-offs among model fidelity, communication efficiency, latency constraints, and fairness under heterogeneous driving conditions. It further discusses governance frameworks, policy implications for data sovereignty, infrastructure requirements for low-latency vehicle-to-everything connectivity, and sustainability considerations. Through analysis of current federated learning protocols and world model architectures, we argue that federated world models offer a viable path toward scalable, privacy-compliant, and collaborative autonomous driving, while also highlighting open challenges in adversarial robustness, non-stationary environments, and certification of distributed intelligent systems.