Outlier Detection in Ad Hoc Networks Using Dempster-Shafer Theory

Mobile Ad-hoc NETworks (MANETs) are known to be vulnerable to a variety of attacks due to lack of central authority or fixed network infrastructure. Many security schemes have been proposed to identify misbehaving nodes. Most of these security schemes rely on either a predefined threshold, or a set of well-defined training data to build up the detection mechanism before effectively identifying the malicious peers. However, it is generally difficult to set appropriate thresholds, and collecting training datasets representative of an attack ahead of time is also problematic. We observe that the malicious peers generally demonstrate behavioral patterns different from all the other normal peers, and argue that outlier detection techniques can be used to detect malicious peers in ad hoc networks. A problem with this approach is combining evidence from potentially untrustworthy peers to detect the outliers. In this paper, an outlier detection algorithm is proposed that applies the Dempster-Shafer theory to combine observation results from multiple nodes because it can appropriately reflect uncertainty as well as unreliability of the observations. The simulation results show that the proposed scheme is highly resilient to attackers and it can converge stably to a common outlier view amongst distributed nodes with a limited communication overhead.