Detection of IoT Cyberattacks in Smart Cities using Deep Neural Networks

Abstract

Nowadays, IoT and smart cities are increasingly becoming popular topics among both researchers and practitioners. IoT applications are the main backbone for building a smart city. Many governments use IoT applications to provide better services for their citizens, and other non-governmental organizations also use them to provide better services and products for their customers. Moreover, the day-to-day activities of society and device interactions in a smart city are carried out over IoT applications. Conversely, new and intelligent attacks are greatly increasing due to the behavior of these applications. As a result, security becomes one of the most crucial concerns that need to be addressed. To date, several intrusion detection models have been proposed by several researchers for ensuring the security of IoT devices in the smart city. In this thesis, I proposed deep neural network-based models MLP, LSTM, and GRU for detecting binary and muti-class IoT cyber attacks, using an imbalanced big data set. The most recent datasets, UNSW-NB15 and CICIDS 2017, were used for model training and evaluation, which are enhanced by a variety of recently added cyber attacks. The experimental results for the UNSW-NB15 dataset show that the MLP model outperformed other models in terms of recall, precision, F1-score, and FPR (false positive rate) with values of 99.17%, 99.17%, 99.17%, and 0.0037, respectively. Furthermore, the LSTM model achieved a higher accuracy of 99.26%. In the case of conventional and ensemble models, Random Forest outclasses other models with respect to all metrics when trained and evaluated with the UNSW-NB15 dataset. Further, when the dataset CICIDS2017 was used for training and evaluating the Random Forest model, it outperformed other conventional and ensemble methods. Among the deep neural network models, the MLP model classified attacks with the accuracy of 98.10%, precision of 98.20%, F1-score of 98.12%, and FPR of 0.0202, which makes it the best-performing deep learning model.