PRISM Policy-Reinforced Lightweight S plit-Federated Learning for Multi-Modal Beamforming in IoV

Multi-modal beamforming in vehicular networks leverages GPS, LiDAR, and camera data to enable faster and more accurate beam alignment, while preserving privacy via Federated Learning (FL). However, existing fusion networks are often large and resource-intensive, and the heterogeneity of vehi cles intensifies computing constraints, causing latency, straggler effects, and high energy consumption in real driving cycles. This paper proposes PRISM, a policy-reinforced lightweight split federated learning framework tailored for multi-modal beam forming in Internet of Vehicles (IoV). To efficiently utilize limited mobile edge computing distributed units (MEC-DUs) in dynamic vehicular environments, PRISM introduces a novel Multi-agent reinforcement learning (MARL) module, Tridec-Qmix, which jointly determines client participation, server assignment, and multi-split-points selection, formulated as a cross-layer multi objective optimization balancing energy efficiency, latency, and network utility. We implement PRISM in a realistic vehicular edge environment with SUMO and NS-3. Results show that PRISM reduces energy consumption by 67.60% and training cost by up to 45.57% compared to state-of-the-art baselines. For inference, PRISM further employs multi-feature matching knowledge distillation (KD) to compress transformer-based fu sion models into lightweight CNNs, shrinking model size by over 50% to 15.82 MB while maintaining accuracy.