Abstract
Intrusion detection systems are one of the security tools widely deployed in network architectures in order to monitor, detect and eventually respond to any suspicious activity in the network. However, the constantly growing complexity of networks and the virulence of new attacks require more adaptive approaches for optimal responses. In this work, we propose a semi-supervised approach for network anomaly detection inspired from the biological negative selection process. Based on a reduced dataset with a filter/ranking feature selection technique, our algorithm, namely negative selection for network anomaly detection (NSNAD), generates a set of detectors and uses them to classify events as anomaly. Otherwise, they are matched against an Artificial Human Leukocyte Antigen in order to be classified as normal. The accuracy and the computational time of NSNAD are tested under three intrusion detection datasets: NSL-KDD, Kyoto2006+ and UNSW-NB15. We compare the performance of NSNAD against a fully supervised algorithm (Naive Bayes), an unsupervised clustering algorithm (K-means) and a semi-supervised algorithm (One-class SVM) with respect to multiple accuracy metrics. We also compare the time incurred by each algorithm in training and classification stages.