Joint Power Allocation and Accuracy Optimization for Federated Learning in ISAC-Enabled Vehicular Networks with NOMA

Abstract

Vehicular networks supporting autonomous driving must simultaneously perform federated learning (FL) for distributed model training and radar sensing for environmental perception. Both functions share the same onboard hardware and transmit power budget, which imposes a trade-off between radar sensing performance and FL model upload efficiency. The existing works on FL in vehicular networks ignore sensing constraints, while integrated sensing and communication (ISAC) studies do not account for the FL training. To bridge this gap, this paper investigates a unified FL framework for ISAC-enabled vehicular networks, where non-orthogonal multiple access (NOMA) is adopted to particularly support simultaneous multi-vehicle model uploads over the same frequency band. We formulate a latency minimization problem by jointly optimizing local computation accuracy and transmit power allocation for both radar and communication, subject to the corresponding communication and sensing constraints. The optimization problem is non-convex due to the coupling among variables and system requirements. To solve it, we develop an efficient iterative algorithm based on successive convex approximation and alternating optimization. Simulation results demonstrate the effectiveness of the proposed approach in reducing FL training latency compared to baseline schemes, while ensuring reliable sensing performance.