A Semantic Communication-Assisted Federated Learning Framework for Internet of Vehicles

Federated learning (FL) has wide applications in the internet of vehicles (IoV), but it requires vehicles to perform local training and upload complete model parameters, resulting in heavy communication overhead. To address this issue, this paper proposes a semantic communication-assisted FL (SeFL), which reduces communication overhead by uploading semantic features instead of model parameters. Particularly, the vehicle side employs pre-trained encoders to extract and transmit features, while the server aggregates features for centralized task head training. Moreover, SeFL mitigates the negative impact of non-independent and identically distributed (non-IID) data on model performance through the server centralized training, which aggregates multi-source semantic features to reconstruct global data distribution. Besides, SeFL treats delayed semantic features as new samples participating in iterations, effectively avoiding convergence instability caused by outdated parameter aggregation. Experimental results demonstrate that when using ResNet-101 as the backbone network with a per-round aggregation time threshold (Tthreshold) of 1.0 s, SeFL achieves classification accuracy improvements of about 8.8% and 18.0% compared to traditional FL (TFL) and hierarchical FL (HFL), respectively. SeFL also reduces total communication overhead by about 99.2% and 98.5% compared to TFL and HFL, respectively. It also shows that when the Tthreshold is decreased from 1.0 s to 0.3 s, SeFL maintains relatively stable accuracy while TFL accuracy drops by approximately 28.9%.