Secure cluster-based hierarchical federated learning in vehicular networks

Abstract

Hierarchical Federated Learning (HFL) improves scalability in vehicular networks by aggregating local updates through intermediate cluster heads (CHs). Yet, this hierarchy can amplify the impact of poisoned or unreliable vehicles and degrade model integrity. Most existing poisoning defenses for federated learning rely on fixed filtering and do not exploit hierarchical structure, cross-cluster validation, or reliability-aware participation in mobility-driven settings with non-independent and identically distributed (non-IID) vehicular data. To address these gaps, we introduce DARCS, a secure cluster-based HFL framework that integrates anomaly detection and reliability-based client selection within the HFL aggregation process. Vehicles are evaluated using a reliability score built from historical accuracy, contribution frequency, and anomaly records. DARCS applies Z-score analysis on update norms to expose Gaussian noise poisoning, and uses cosine similarity to capture directional inconsistencies typical of gradient ascent attacks. It also refines screening via an accuracy-driven adaptive thresholding mechanism and evolved packet core (EPC)-level cross-cluster validation. Reliability-weighted aggregation further down-weights low-trust contributions without introducing additional communication rounds beyond baseline cluster-based HFL. Extensive simulations demonstrate that DARCS substantially improves stability and convergence, keeping task accuracy within about 2–3% of the attack-free reference while reducing convergence time by up to 17.1% compared with cosine-similarity, Z-score, and their combination defenses.