Unsupervised anomaly detection for unlabelled wireless sensor networks data

Abstract

With the advances in sensor technology, sensor nodes, the tiny yet powerful device are used to collect data from the various domain. As the sensor nodes communicate continuously from the target areas to base station, hundreds of thousands of data are collected to be used for the decision making. Unfortunately, the big amount of unlabeled data collected and stored at the base station. In most cases, data are not reliable due to several reasons. Therefore, this paper will use the unsupervised one-class SVM (OCSVM) to build the anomaly detection schemes for better decision making. Unsupervised OCSVM is preferable to be used in WSNs domain due to the one class of data training is used to build normal reference model. Furthermore, the dimension reduction is used to minimize the resources usage due to resource constraint incurred in WSNs domain. Therefore one of the OCSVM variants namely Centered Hyper-ellipsoidal Support Vector Machine (CESVM) is used as classifier while Candid-Covariance Free Incremental Principal Component Analysis (CCIPCA) algorithm is served as dimension reduction for proposed anomaly detection scheme. Environmental dataset collected from available WSNs data is used to evaluate the performance measures of the proposed scheme. As the results, the proposed scheme shows comparable results for all datasets in term of detection rate, detection accuracy and false alarm rate as compared with other related methods.