Selective Features Based Machine Learned Intrusion Detection Framework for Wireless Sensor Networks: A Need for Cryptographic Approach

Abstract

Intruder attacks are a curse to wireless sensor network (WSN) nodes that affect their regular performance differently. The attacker cheats the cooperative characteristics of sensing elements pretending as a trustworthy node. Such intruders drastically degrade the WSN performance by affecting the sensing, collecting, processing, and transmitting capabilities of ideal nodes in the network. Due to their pretending behaviour, detecting such harmful nodes in the network is difficult since they are sometimes foes and other times friends. This paper introduces a selective feature-based machine-learning (SF-BML) framework that can detect an intruder node with a higher accuracy. The WSN-DS dataset constructed for four different attacks (Flooding, Blackhole, Grayhole, and Scheduling) is experimented with different feature attributes. The optimum features are found and tested for detection accuracy relating to all five categories.  The experimental results using six significant features (SF) and a combination of eight other semi-significant features (SSF) showed that the normal-attacker node's detection accuracy increases with increasing features up to fourteen. The maximum training and test accuracy using fourteen features for the support vector machine (SVM) was 99.17% and 99.00% respectively.