Learning to Transfer Knowledge Between Datasets to Enhance Intrusion Detection Systems

Abstract

Software-defined network (SDN) is a technology that is being used widely to reduce the time and effort required for programming network functions. However, by splitting the control layer and data layer, the SDN architecture also attracts numerous types of attacks such as spoofing or information disclosure. In the recent years, a few research articles coped with the security problem by introducing open datasets and classification techniques to detect the attacks to SDN. The state-of-the-art techniques perform very well in a single cross-validation dataset, i.e., in the situation, the training and the evaluation datasets are being withdrawn from the same source. However, their performance reduces significantly in the presence of concept drift, i.e., if the testing dataset is collected from a different source than the observed dataset. In this research study, we address this cross-dataset predictive issue by several concept drift detection techniques. The experimental results let us claim that our presented models can improve the performance in the cross-dataset scenario.