Abstract
In recent times, there has been an extensive and expeditious growth and advancement of information and communication related technologies as well an advancement of the Internet. With these technological breakthroughs came the rapid development of wireless enabled devices. Consequently, the expansion of wireless networks capable of handling the increasing volume of information generated by those networks became inevitable. As a result, wireless networks are vulnerable and exposed to numerous security threats as well as privacy concerns. Currently, the existing protective and preventive measures such as wired and wireless Intrusion Detection Systems (IDSs) are not fully immune to the growing number of network intrusions instances. An IDS has a critical role in ensuring that various networks are secured and protected against attacks. Research has demonstrated that the majority of current IDS systems do not perform at the required level. There exists many types of IDS systems; however, we focused on Machine Learning (ML) and Deep Learning (DL) based IDSs. The performance of current ML and DL based IDS systems for wired and wireless networks suffer from a low level of detection accuracy and a high ratio of false alarm rate. Moreover, the increase in the amount of data generated by the wired and wireless networks has caused the datasets required to design and implement ML and DL based IDSs to become highly dimensional in terms of features and extremely complex in terms of the types of data. In this thesis, we design and implement DL based IDS systems using Feed Forward Deep Neural Networks (FFDNNs), Deep Long-Short Term Memory Recurrent Neural Networks (DLSTM RNNs) and Deep Gated Recurrent Unit Recurrent Recurrent Neural Networks (DGRU RNNs). In the aim to tackle the issue of the highly dimensional input spaces, we further implement an Information Gain (IG) based feature extraction method that is conjoined with the FFDNNs. We also devised and implemented two wrapper-based feature selection algorithms. One is based on the Extra-Trees (ET) classifier and the other is inspired from the Random Forest (RF) classifier. The ET is coupled with the DLSTM RNNs and the DGRU RNNs. The RF is used in conjunction with FFDNNs. In order to evaluate the performance of our frameworks, the following three datasets were used: the NSL-Knowledge Discovery and Data mining (NSL-KDD) dataset, the University of New South Wales-NB15 (UNSW-NB15) dataset and the Aegean Wi-Fi Intrusion Dataset (AWID).
Ph.D. (Electrical and Electronic Engineering)