Distributed anomaly detection scheme based on lightweight data aggregation in wireless sensor network

Abstract

Wireless Sensor Networks (WSNs) have been used in many domains for instance in business applications, industrial applications, and military applications to monitor a phenomenon, track an object, or control a process. As the sensor nodes communicate continuously from the target phenomenon to the base station, hundreds of thousands of multivariate data are collected from sensor nodes will be analysed at the endpoint called base station or sink node for decision making. Unfortunately, data is not usually accurate and reliable which will affect the decision making at the base station. There are many reasons that cause inaccurate and unreliable data such as malicious attack, harsh environment as well as sensor node failure. In the worst-case scenario, the node failure will also lead to the dysfunction of the entire network. Thus, anomaly detection is used to ensure that the data acquired at the endpoint is accurate. On the other hand, as sensor nodes possess resources constraint in terms of energy, processing, and storage, therefore, anomaly detection techniques must be designed in a lightweight manner. Meanwhile, existing anomaly detection techniques pose weaknesses as these have high computational and communication cost, ignore multivariate data and features’ correlation, and some are parameter-dependent. The purpose of this research is to design and develop an efficient and effective anomaly detection scheme for WSN by minimizing the resource constraint in WSNs. This purpose can be achieved by first, applying the feature selection method to select significant features for minimizing the resource utilization. Second, designing an efficient network structure of WSNs architecture based on a data aggregation scheme for reducing data transmission in the network. Third, designing lightweight anomaly detection scheme (CESVM-DR) using One-class Support Vector Machine (OCSVM) anomaly detection scheme and incorporating dimension reduction technique based on Candid Covariance-Free Incremental Principal Component Analysis (CCIPCA) to minimise the computational complexity of covariance matrix in CESVM. Lastly, enhancing the efficiency and effectiveness of the anomaly detection scheme by designing the distributed anomaly detection scheme (DCESVM-DR). The effectiveness and efficiency of the proposed anomaly detection schemes were tested using real-world datasets as well as soil data collected from the palm oil plantation. The results show the proposed CESVM-DR anomaly detection scheme with an average of 92%–100% detection accuracy using the real datasets while minimizing the computational complexity and energy overhead. Meanwhile, exploiting the correlation between sensor nodes to detect the anomalies on the DCESVM-DR has enhanced the effectiveness results as well as minimised the memory complexity and energy.