Performance Evaluation of Unsupervised Learning Techniques for Intrusion Detection in Mobile Ad Hoc Networks

Abstract

Mobile ad hoc network (MANET) is vulnerable to numerous attacks due to its intrinsic characteristics such as the lack of fixed infrastructure, limited bandwidth and battery power, and dynamic topology. Recently, several unsupervised machine-learning detection techniques have been proposed for anomaly detection in MANETs. As the number of these detection techniques continues to grow, there is a lack of evidence to support the use of one unsupervised detection algorithm over the others. In this paper, we demonstrate a research effort to evaluate the effectiveness and efficiency of different unsupervised detection techniques. Different types of experiments were conducted, with each experiment involves different parameters such as number of nodes, speed, pause time, among others. The results indicate that K-means and C-means deliver the best performance overall. On the other hand, K-means requires the least resource usage while C-means requires the most resource usage among all algorithms being evaluated. The proposed evaluation methodology provides empirical evidence on the choice of unsupervised learning algorithms, and could shed light on the future development of novel intrusion detection techniques for MANETs.