FedDLD: Dual-Level Federated Distillation With Adaptive Knowledge Transfer for DAG-Secured IoVs
Sa Xiao, Xiaoge Huang, Mu Zhou, Chengchao Liang, Qianbin Chen · IEEE Transactions on Vehicular Technology · 2025
While federated learning (FL) offers privacy-aware collaborative intelligence for Internet of Vehicles (IoV), existing frameworks struggle with communication efficiency and adaptability under spatiotemporal data heterogeneity in mobile environments. This paper proposes a Federated Dual-Level Distillation (FedDLD) framework, which integrates hierarchical knowledge transfer with secure aggregation mechanisms. Two coordinated tiers are considered: edge-anchored federated distillation ensures essential model performance, followed by Dynamic Adaptive Mutual Distillation (DAMD) improves knowledge transfer efficiency among connected autonomous vehicles (CAVs) while reducing communication overhead. Moreover, a Directed Acyclic Graphs (DAG) based blockchain layer is embedded to guarantee secure decentralized model sharing with lightweight computation overhead. Furthermore, theoretical analysis proves the FedDLD's convergence and boundedness. Experimental evaluations on CIFAR-10 and traffic sign datasets show that the proposed FedDLD achieves 12.3% higher accuracy and 35.7% lower communication overhead compared to state-of-the-art methods, while maintaining robustness against label flipping attacks (LF) in dynamic IoV environments.