Data-driven Extreme Events Modeling for Vehicle Networks by Personalized Federated Learning: Invited Paper

Managing queuing latency is crucial to achieve ultra-reliable low-latency communications (URLLC) in the future vehicle networks. In this work, we propose a novel joint power and resource allocation strategy to enhance the worst-case relia-bility by minimizing the network-wide maximum queue length. A constraint of a long-term energy budget is considered, as vehicles must simultaneously ensure other tasks. In addition, vehicle communications are assumed to have a heterogeneous nature and the distribution of extreme events may vary between vehicles, while in this work extreme value theory (EVT) is exploited to model these extreme events. Moreover, personalized federated learning is employed to learn the distribution while handling the heterogeneity among vehicles. Simulation results confirm that the proposed design reduces the length of the worst-case queuing latency and that, in comparison to traditional federated learning, the introduced personalized federated learning approach signif-icantly increases the estimation accuracy of local extreme event distribution without increasing the communication load.