Detection of IoT Cyberattacks in Smart Cities: A Comparative Analysis of Deep Learning and Ensemble Learning Methods

Abstract

In this study, we embarked on a comparative investigation of Deep Learning (DL) techniques and ensemble learning approaches for enhancing IoT security. Specifically, we scrutinized the performance of Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Random Forest (RF), and AdaBoost models evaluated to both binary and specific attack vector classifications. The imbalanced and voluminous datasets of UNSW-NB15 and CICIDS2017 were employed for evaluation. The empirical evidence gleaned from our experiments suggests that RF exhibits superior efficacy over its counterparts, with accuracy and F1-score in the range of 99.68% to 99.90%. Within the DL paradigm, the MLP model achieved the highest F1-score (99.17%) and the lowest False Positive Rate (FPR) of 0.0037 using UNSW-NB15, among DL models. Overall, the proposed models exhibit commendable performance in binary classification tasks. However, this does not indicate their suitability for the detection of all types of attacks, as the individual attack detection result shows. Furthermore, models employed in our work demonstrated superior results as compared to existing models that used smaller sample sizes of these datasets.