FedVR360: Federated learning enabled privacy-preservation for VR 360° video streaming in vehicular edge computing

The integration of Virtual Reality (VR) 360° video streaming into Vehicular Edge Computing (VEC) enables immersive in-vehicle experiences but introduces significant challenges in privacy preservation and real-time performance. Existing predictive caching solutions rely on centralized learning, which requires aggregating sensitive user data such as head orientation and vehicle trajectories, thereby violating privacy regulations and exposing users to inference attacks. Federated Learning (FL) offers a promising solution; however, its deployment for VR streaming in VEC is constrained by highly non-Independent and Identically Distributed (non-IID) data, intermittent connectivity, and the need for joint multi-modal prediction without raw data exchange. To address these limitations, this paper proposes FedVR360, a comprehensive privacy-preserving FL framework for joint trajectory and viewport prediction in vehicular edge environments. FedVR360 integrates a federated multi-modal Temporal Fusion Transformer with prototype-based cross-modal fusion, asynchronous hierarchical aggregation across vehicles and roadside units, and provides formal privacy guarantees using Rényi Differential Privacy. Additionally, a hybrid personalization strategy mitigates non-IID degradation. The performance evaluations conducted on real VR viewport traces and simulated vehicular trajectories show that FedVR360 closely approaches centralized performance, recovering over 84% of the centralized trajectory prediction gap while achieving viewport prediction accuracy (F1@10 = 0.795) that surpasses all privacy-preserving federated baselines under the same privacy budget. Under a practical privacy budget, FedVR360 reduces membership inference attack success to near random guessing (50.5%) and significantly degrades gradient inversion attacks. FedVR360 achieves corresponding improvements in precision and recall across all evaluation metrics, reduces normalized trajectory prediction error (MAE = 0.282), and maintains per-client prediction variance below 0.15 under non-IID data. Additionally, the framework ensures real-time inference latency below 30 ms with moderate training and memory overhead, demonstrating a balanced trade-off between accuracy, efficiency, and privacy under controlled simulation conditions.