Mobility-Aware Predictive Split Federated Learning for 6G vehicular networks with ultra-low latency guarantees

The integration of distributed learning in 6G vehicular networks faces significant challenges due to high mobility, stringent latency requirements, and resource constraints at the network edge. This paper proposes MAPSFL, a novel mobility-aware predictive split federated learning framework that seamlessly integrates mobility prediction, dynamic model splitting, and hierarchical learning architectures to enable efficient distributed learning in highly mobile vehicular environments. Our framework employs a predictive mobility model to optimize resource allocation and model splitting decisions while maintaining ultra-low latency guarantees through adaptive compression and selective parameter transmission mechanisms. Theoretical analysis provides convergence guarantees under dynamic network conditions, while extensive experimental results demonstrate that MAPSFL achieves 31% reduction in CPU utilization, 28% lower bandwidth consumption, and 34% reduction in end-to-end training latency compared to state-of-the-art approaches. The proposed work achieved 85% efficiency at large scales of vehicles, i.e. 5000, while ensuring the required latency of 100ms, thus making it particularly suitable for safety-critical vehicular applications. The comprehensive evaluation of the proposed method validates its effectiveness in addressing the challenges of high mobility, resource constraints, and network dynamics in 6G vehicular networks.