A hybrid data mining anomaly detection technique in ad hoc networks. (30th May 2007)

• Record Type: Journal Article
• Authors: Liu, Yu; Li, Yang; Man, Hong; Jiang, Wei
• Is Part Of: International journal of wireless and mobile computing. Volume 2:Number 1(2007)
• Page Start: 37
• Page End: 46
• Publication Date: 2007-05-30
• Subjects: anomaly detection -- ad hoc networks -- association rules -- cross-features -- data mining -- Bayesian networks -- mobile networks -- wireless networks -- network security -- intrusion detection -- simulation; Mobile computing -- Periodicals; Wireless communication systems -- Periodicals; 004.6
• Languages: English
• ISSNs: 1741-1084

Abstract

Ad hoc network security mainly relies on defence mechanisms at each mobile node due to lack of infrastructure. For this reason, various intrusion detection techniques have been proposed for ad hoc networks. Developing Intrusion Detection Systems (IDS) for individual nodes in ad hoc network is challenging for a number of reasons, including resource constraints at each node and the difficulties to locate attack source for prompt response. In this paper, we propose a hybrid data mining anomaly detection technique for node-based IDS. Specifically, we incorporate two data mining techniques, that is, association-rule mining and cross-feature mining, to characterise normal behaviours of mobile nodes and detect anomalies by finding deviance from the norm. The advantage of our hybrid approach is that association-rule mining and cross-feature mining usually complement each other in time scale and sensitivity to different attack types. We investigate features of interest from both the medium access (MAC) layer and the network layer. Our intention of using the MAC layer features is to localise the attack source within one-hop perimeter. To preserve the precious energy of mobile nodes, we propose two compact feature sets, that is, direct feature set and statistical feature set, that target on short-term and long-term profiling of normal node behaviours, respectively. Considering the characteristic of audit data collected upon different feature sets, we apply association-rule mining to the short-term profiling and cross-feature mining to the long-term profiling. We validate our work through ns-2 simulation experiments. Experimental results show the effectiveness of our method.