A Comparative Study of Machine Learning Models for Cyber-attacks Detection in Wireless Sensor Networks

Abstract

Wireless Sensor Network is a key technology for Internet of Things, but it is an example of energy-restricted networks. In such networks with a large number of deployed sensors with limited capabilities, security becomes a crucial concern which requires proper design of detection and mitigation techniques that address such challenge. Machine learning is one of the most efficient ways to design cyber-attacks detection systems for WSN. This paper presents a comparative study and performance analysis of different machine learning classification techniques for cyber-attacks detection in WSN. We investigate the performance of three recent boosting techniques, namely GBM, LightGBM, and Catboost. Their performances are compared to those of three commonly known machine learning methods, namely Gaussian NB, KNN, and RF. We apply both Pearson’s correlation and mutual information as an ensemble feature selection and dimensionality reduction. Tree-based Pipeline Optimization Tool is utilized for hyperparameter tuning. We use the specialized dataset, WSN-DS, that contains numerous samples corresponding to four kinds of attacks: Grayhole, Blackhole, Flooding, and TDMA scheduling. The performance of the six models is investigated in terms of accuracy, probability of detection, probability of false alarm, probability of misdetection, memory usage, processing time, and average prediction time per sample.