CoSen-IDS: A Novel Cost-Sensitive Intrusion Detection System on Imbalanced Data in 5G Networks

Lu Yuan, Jiyan Sun, Shangyuan Zhuang, Yinlong Liu, Liru Geng, Wei Ma

Abstract: With the widespread deployment of 5G networks, there is a massive increase in data volume and types of 5G network traffic. This results in the emergence of severely imbalanced data distribution, including imbalances between benign and malicious traffic, as well as imbalances among different types of malicious traffic. This imbalanced data distribution presents significant challenges for intrusion detection systems (IDSs). Existing IDSs in 5G networks often treat diverse traffic classes equally, overlooking the varying costs of misclassification across classes. This leads to lower accuracy for minority traffic classes. Additionally, researchers use various sampling methods to balance benign and malicious traffic in 5G networks but often ignore imbalances among different types of malicious traffic, hindering specific identification improvements. Moreover, the redundant features employed by current IDSs contribute to suboptimal detection performance across varied scenarios in 5G networks. In this paper, we present CoSen-IDS, an innovative intrusion detection system designed for 5G networks based on cost-sensitive learning to solve imbalanced issues. Specifically, CoSen-IDS dynamically identifies the crucial features within 5G network traffic by evaluating feature importance. Moreover, we propose a cost-sensitive strategy that guides Generative Adversarial Networks to amplify multiple minority traffic classes simultaneously in the training dataset. Finally, we aggregate different base classifiers to attain a more robust intrusion detection model. Through extensive experiments on two datasets, our results show that CoSen-IDS outperforms relevant state-of-the-art methods.