Two-Stage Federated Learning Strategy for Fairness and Security in Vehicular Networks

In vehicular networks, securing the network relies on analyzing extensive vehicle perception data to create an accurate road information model. Effective data sharing is crucial for this objective. However, traditional centralized data sharing methods jeopardize the privacy and security of data providers. Federated Learning has emerged as a solution to this challenge. Unfortunately, conventional Federated Learning strategies often overlook the differences in resource investments and data quality among participating vehicles, leading to unfair model distribution and diminishing the willingness of vehicles to share data. This paper proposes a two-stage Federated Learning strategy. In the first phase, aggregation nodes employ a fuzzy comprehensive evaluation method to assess the training results of participating vehicles and allocate aggregation weights accordingly. In the second phase, aggregation nodes calculate vehicle reputations using a periodic averaging method and distribute a differentially compressed global model based on these reputation values. Experimental results demonstrate that the proposed strategy achieves performance comparable to existing Federated Learning algorithms while effectively calculating vehicle reputations, ensuring fair distribution, and mitigating malicious clients.