Deep Learning-Based Anomaly Detection in Connected Vehicle Networks: A Systematic Survey

With the rise of connected vehicles, intelligent transportation systems have revolutionized, rendering them more vulnerable to cyber-attacks, sensor failures, and network intrusions. This systematic survey, conducted in accordance with the PRISMA protocol, synthesizes ۸۵ peer-reviewed articles published between ۲۰۱۰ and ۲۰۲۵ that examine deep learning architectures for anomaly detection. Sources such as Web of Science, Scopus, and IEEE Xplore, among others, were evaluated to categorize deep learning models, including CNNs, LSTMs, and autoencoders, based on supervised, unsupervised, and hybrid approaches. The paper assesses the collective performance of these models for intrusion detection, fault detection, and identification, as well as behavioral anomalies. According to the findings in this research, unsupervised models ۴۵% perform best and produce AUC scores as high as ۰.۹۹ despite sim-to-real differences and asset limitations. As a result, a framework that combines edge computing and ۵G/۶G is advantageous for implementation, as the unsupervised model has exhibited an inability to scale. As a result of the report, the performance of DL was shown to outperform traditional approaches in the dynamic scenario of V۲X data management. It highlighted the significance of consistent dataset sizes and quality, as well as federated learning approaches to improve generalization and privacy in ITS.

کلیدواژه ها:

Deep Learning ، Anomaly Detection ، Connected Vehicles ، Vehicular Networks