Anomaly detection for IoT networks using machine learning

Abstract

The Internet of Things (IoT) is considered one of the trending technologies today. IoT affects various industries, including logistics tracking, healthcare, automotive and smart cities. A rising number of cyber-attacks and breaches are rapidly targeting networks equipped with IoT devices. This thesis aims to improve security in IoT networks by enhancing anomaly detection using machine learning. This thesis identified the challenges and gaps related to securing the Internet of Things networks. The challenges are network size, the number of devices, the human factor, and the complexity of IoT networks. The gaps identified include the lack of research on signature-based intrusion detection systems used for anomaly detection, in addition to the lack of modelling input parameters required for anomaly detection in IoT networks. Furthermore, there is a lack of comparison of the performance of machine learning algorithms on standard and real IoT datasets. This thesis creates a dataset to test the anomaly binary classification performance of the Neural Networks, Gaussian Naive Bayes, Support Vector Machine, and Decision Trees machine learning algorithms and compares their results with the KDDCUP99 dataset. The results show that Support Vector Machine and Gaussian Naive Bayes perform lower than the other models on the created IoT dataset. This thesis reduces the number of features required by machine learning algorithms for anomaly detection in the IoT networks to five features only, which resulted in reduced execution time by an average of 58%. This thesis tests CNNwGFC, which is an enhanced Convolutional Neural Network model, in detecting and classifying anomalies in IoT networks. This model achieves an increase of 15.34% in the accuracy for IoT anomaly classification in the UNSW-NB15 compared to the classic Convolutional Neural Network. The CNNwGFC multi-classification accuracy (96.24%) is higher by 7.16 than the highest from the literature.