Optimal intrusion detection for imbalanced data using Bagging method with deep neural network optimized by flower pollination algorithm

As the number of connected devices and Internet of Things (IoT) devices grows, it is becoming more and more important to develop efficient security mechanisms to manage risks and vulnerabilities in IoT networks. Intrusion detection systems (IDSs) have been developed and implemented in IoT networks to discern between regular network traffic and potential malicious attacks. This article proposes a new IDS based on a hybrid method of metaheuristic and deep learning techniques, namely, the flower pollination algorithm (FPA) and deep neural network (DNN), with an ensemble learning paradigm. To handle the problem of imbalance class distribution in intrusion datasets, a roughly-balanced (RB) Bagging strategy is utilized, where DNN models trained by FPA on a cost-sensitive fitness function are used as base learners. The RB Bagging strategy derives multiple RB training subsets from the original dataset and proper class weights are incorporated into the fitness function to attain unbiased DNN models. The performance of our IDS is evaluated using four commonly utilized public datasets, NSL-KDD, UNSW NB-15, CIC-IDS-2017, and BoT-IoT, in terms of different metrics, <i>i.e</i>., accuracy, precision, recall, and F1-score. The results demonstrate that our IDS outperforms existing ones in accurately detecting network intrusions with effective handling of class imbalance problem.