SVQ-VAE: Federated Learning-Based Semantic-Aware Communication for Vehicular Networks

The integration of semantic communication technology into intelligent vehicular networks represents a promising research direction, as it significantly reduces data transmission volume and spectrum usage, addressing the high demands of transmitting large-scale visual information between vehicles. Existing studies typically assume that communicating parties share a common database. However, this assumption poses significant privacy risks, particularly in inter-vehicle scenarios. To address these challenges, we propose a federated learning-based semantic-aware communication for vehicular networks. In this approach, each intelligent vehicle locally trains a semantic communication model and uploads its model parameters to the edge server for aggregation. To further enhance data transmission efficiency, we introduce a semantic-aware vector quantized variational autoencoder (SVQ-VAE) architecture as the local semantic communication model. This architecture optimizes transmission by selectively compressing and quantizing only the most relevant semantic information for the task. Additionally, to address data heterogeneity among vehicles, we propose a hypernetwork-based personalized federated learning (HPFL) scheme. This approach enhances the model’s scalability and generalization by training a hypernetwork at the edge server to generate specific weight parameters for each vehicle’s semantic communication model. Simulation experiments on the CIFAR-10 and BDD100K datasets demonstrate that our proposed federated learning-based semantic-aware communication achieves superior task completion rates, semantic transmission efficiency and transmission delay compared to existing federated semantic communication architectures.