An Efficient Anomaly Detection System in IoT Edge using Chi Square-Improved Particle Swarm Optimization Feature Selection with Ensemble classifiers

Abstract

Anomaly detection using machine learning (ML) algorithms is the key research theme in the modern digital era. Though the recent ML-based anomaly detection models have better detecting ability, the vast volume of data and its multi-dimensionality limit their ability with less accuracy, a low detection rate, and high learning complexity. This paper aims to enhance the performance of anomaly detection by combining various optimized ensemble learning algorithms, such as random forest (RF), extreme gradient boosting (XG Boost), adaptive boosting (Ada Boost), and light gradient boosting machine (LGBM), with a new hybrid feature selection approach. An evolved version of particle swarm optimization (IPSO) is initially developed, which integrates the elimination and opposition-based learning approaches to enhance PSO and then hybridizes it with the Chi-square method (Chi-IPSO). The developed model is evaluated using two standard datasets: UNSW NB 15, and CICIDS 2017. The research results show that the RF algorithm with Chi-IPSO performs better with an accuracy of 94.58% for the UNSW NB 15, and 99.70% for the CICIDS 2017. Several assessment measures, including F-score, MCC value, accuracy, precision, and recall, are used to highlight the outcome analysis of the suggested model. The results clearly show that the created model performs better than other modern approaches.