Mobility-Aware Cross-Region Federated Mutual Knowledge Distillation in Internet of Vehicles

By collaboratively training a global model without accessing private data from connected and autonomous vehicles (CAVs), federated learning (FL) enables efficient and privacy-preserving sharing and utilization of data in the internet of vehicles. However, in the real scenarios, these data often presents significant non-independent and identically distributed characteristics, which will affect model performance and communication efficiency. To address these problems, in this paper, we propose a mobility-based cross-region dynamic federated mutual knowledge distillation (MC-MKD) method. The proposed MC-MKD method first introduces a mobility-based CAV selection method and gives priority to CAVs with longer residence time and higher data quality, which ensures the collected data better represent the global data distribution. Secondly, a data filtering strategy based on density ratio estimation is proposed to eliminate abnormal samples with large differences in distribution from the perception range of the roadside unit. Building on this, the MC-MKD method employs a federated distillation (FD) framework with privacy protection, incorporates an R-Dropout regularization mechanism to mitigate overfitting, and introduces dynamic bidirectional knowledge distillation to achieve complementary knowledge transfer between the teacher and the student models. Experimental results show that the proposed MC-MKD method is superior to the existing FL methods in terms of model accuracy, communication overhead control, and privacy protection.