Environment-Aware Personalized Heterogeneous Federated Distillation for Dual-Layer Blockchain-Enabled Internet of Vehicles

In the Internet of Vehicles (IoV), personalized federated learning (PFL) can generate personalized models for diverse local data distributions while preserving user privacy. However, previous PFL approaches often apply weighted aggregation based on the model parameters or loss values, ignoring the layer-level differences of models and the skewed distribution of classes. Additionally, their performance in heterogeneous systems and personalized services is severely limited by the same model structure. Knowledge distillation (KD) can enable heterogeneous model collaboration, but its performance is usually constrained by the quality of public data. To address these issues, in this paper, we propose a dual-layer blockchain-enabled environment-aware personalized federated distillation (DB-EPFD) approach in IoV, which simultaneously improves system security, model accuracy, and vehicle fairness. Firstly, vehicles are grouped by perceiving environmental characteristics to mitigate statistical heterogeneity, while their local model structures can be customized to deeply match heterogeneous resources. Moreover, a dual-layer blockchain is introduced to enable intra-independent and inter-collaborative secure data sharing in asynchronous PFL. For intra-group PFL, a layer-wise adaptive model aggregation (LAMA) algorithm is proposed to acquire precise aggregation weights and high vehicle fairness. For inter-group PFL, a multi-dimensional knowledge-fused federated distillation (MKFD) algorithm is designed to enable high-quality KD on heterogeneous models in a data-free manner. Finally, simulation results demonstrate the superiority of the proposed approach in highly heterogeneous scenarios.