Detection of Attacks in Wireless Sensor Networks Using Unsupervised Learning Approach

Abstract

This paper presents an unsupervised machine learning based attack detection approach for wireless sensor networks (WSNs). Two unsupervised methods, namely Fuzzy-c-means (FCM) and K--means clustering, are employed. Principal Component Analysis (PCA) is implemented in both methods as a feature selection technique. The models are trained and tested on the WSN-DS and KDD Cup 99 datasets. These datasets are originally labeled, and by removing the labels they are transformed into unlabeled datasets. The proposed unsupervised learning approach is then applied to these unlabeled datasets, and the number of clusters is finally validated using the original labeled datasets. The attack types considered include blackhole, grayhole, flooding, TDMA, probe, U2R, and R2L.