Scalable hierarchical federated graph-transformer architecture for efficient multi-modal intrusion detection in 6G...

The increasing growth of 6G-empowered UAV-assisted vehicular IoT systems brings forth unprecedented scalability issues for distributed intrusion detection, especially in the context of non-IID data distributions and heterogeneous edge environments. Centralized and flat federated systems do not efficiently coordinate large-scale, latency-sensitive and resource-constrained nodes. In this research, we present a scalable hierarchical federated Graph-Transformer architecture for effective multi-modal intrusion detection spanning UAV-edge-cloud tiers. The platform employs hierarchical aggregation among cars, UAVs, and regional edge servers for reducing communication overhead (CO) and speeding up convergence in non-IID scenarios. Meanwhile, a hybrid Graph Neural Network (GNN) and Transformer backbone is adopted to model spatial topology and temporal dynamics, and a lightweight multi-modal fusion is employed to integrate network traffic, telemetry and channel condition information. To improve the efficiency of the system, we propose adaptive aggregation scheduling and communication compression algorithms that considerably reduce bandwidth consumption and training latency. The experimental results on the CIC-IoT-2023, ToN-IoT and Edge-IIoTset datasets exhibit enhanced scalability with over 98.20% detection accuracy and up to 38.00% transmission cost reduction compared to the flat federation baselines. The work presents a scalable and system-efficient approach for next generation distributed intrusion detection in large-scale 6G vehicle ecosystems.