A Novel Federated Learning Framework Based on Trust Evaluation in Internet of Vehicles
Na Wan and Denghui Wang
With the continuous progress of the society, Intelligent Connected Vehicles (ICVs) in Internet of Vehicles (IoVs) continuously generate a large amount of data. Considering the data security at the vehicle node end, the introduction of federated learning (FL) into the Internet of Vehicles may be an effective solution. However, there are some challenges in applying existing federated learning frameworks directly to IoVs. For example, the participation of malicious nodes in model training can cause federated model poisoning, and the non-IID data generated at the vehicle side can slow down the convergence speed of FL. To address these challenges, this paper proposes a novel federated learning framework based on trust evaluation. The trust evaluation module of this framework evaluates the trust of vehicle nodes from three levels: the success rate of data transmission, the reliability and real-time performance of the network, and the historical behaviors of the nodes to derive the total trust value of the vehicle nodes in order to select nodes with high trust to participate in model training. The model training module of the framework takes into account the problem of slow model convergence caused by the non-IID nature of the training data at the vehicle node side, which is mitigated with the aid of the Federated Adaptive Weighting algorithm (FedAdp). Simulation results on MNIST and Fashion-MNIST datasets demonstrate that the framework proposed in this paper can achieve 85%-90% test accuracy with training data in IID and non-IID settings and two model attack settings.
Keywords: Federated Learning; Trust Evaluation; Internet of Vehicles