Applying Transfer Learning Approaches for Intrusion Detection in Software-Defined Networking

Abstract

In traditional network management, the configuration of routing policies and associated settings on individual routers and switches was performed manually, incurring a considerable cost. By centralizing network management, software-defined networking (SDN) technology has reduced hardware construction costs and increased flexibility. However, this centralized architecture renders information security vulnerable to network attacks, making intrusion detection in the SDN environment crucial. Machine-learning approaches have been widely used for intrusion detection recently. However, critical issues such as unknown attacks, insufficient data, and class imbalance may significantly affect the performance of typical machine learning. We addressed these problems and proposed a transfer-learning method based on the SDN environment. The following experimental results showed that our method outperforms typical machine learning methods. (1) our model achieved a F1-score of 0.71 for anomaly detection for unknown attacks; (2) for small samples, our model achieved a F1-score of 0.98 for anomaly detection and a F1-score of 0.51 for attack types identification; (3) for class imbalance, our model achieved an F1-score of 1.00 for anomaly detection and 0.91 for attack type identification. In addition, our model required 15,230 seconds (4 h 13 m 50 s) for training, ranking second among the six models when considering both performance and efficiency. In future studies, we plan to combine sampling techniques with few-shot learning to improve the performance of minority classes in class imbalance scenarios.