Performance Comparison of IoT Classification Models using Ensemble Stacking and Feature Importance

Abstract

Internet of Things (IoT) security is becoming a top priority as the number of connected devices increases online. This research utilizes the CIC IoT ATTACK 2023 dataset from the University of Brunswick, which consists of 46 million data on various types of attacks on IoT devices, such as DDoS, DoS, Brute Force, Spoofing, and Mirai attacks. To address the imbalance in the dataset, a random undersampling technique is applied to ensure the machine learning model is not biased towards the majority class. The ensemble learning approach was chosen due to its ability to combine the strengths of multiple algorithms, thus improving accuracy and stability in detecting complex IoT attacks. The algorithms used include gradient boosting, bagging, voting, and stacking. In particular, the stacking model, which combines the bagging classifier and gradient boosting, achieved the highest accuracy of 93%. Although the accuracy of the stacking model decreased to 92.4% after feature selection, the precision, recall, and F1-score remained high at 92.0. In addition, the computation time was also reduced from 2111.69 seconds to 1208.27 seconds. These findings indicate that ensemble learning approaches and feature selection techniques have great potential in improving IoT security, providing more reliable and efficient threat detection solutions.