Doctor of Philosophy
With the exponential growth of network-based applications globally, there has been a transformation in organizations' business models. Furthermore, cost reduction of both computational devices and the internet have led people to become more technology dependent. Consequently, due to inordinate use of computer networks, new risks have emerged. Therefore, the process of improving the speed and accuracy of security mechanisms has become crucial.Although abundant new security tools have been developed, the rapid-growth of malicious activities continues to be a pressing issue, as their ever-evolving attacks continue to create severe threats to network security. Classical security techniquesfor instance, firewallsare used as a first line of defense against security problems but remain unable to detect internal intrusions or adequately provide security countermeasures. Thus, network administrators tend to rely predominantly on Intrusion Detection Systems to detect such network intrusive activities. Machine Learning is one of the practical approaches to intrusion detection that learns from data to differentiate between normal and malicious traffic. Although Machine Learning approaches are used frequently, an in-depth analysis of Machine Learning algorithms in the context of intrusion detection has received less attention in the literature.Moreover, adequate datasets are necessary to train and evaluate anomaly-based network intrusion detection systems. There exist a number of such datasetsas DARPA, KDDCUP, and NSL-KDDthat have been widely adopted by researchers to train and evaluate the performance of their proposed intrusion detection approaches. Based on several studies, many such datasets are outworn and unreliable to use. Furthermore, some of these datasets suffer from a lack of traffic diversity and volumes, do not cover the variety of attacks, have anonymized packet information and payload that cannot reflect the current trends, or lack feature set and metadata.This thesis provides a comprehensive analysis of some of the existing Machine Learning approaches for identifying network intrusions. Specifically, it analyzes the algorithms along various dimensionsnamely, feature selection, sensitivity to the hyper-parameter selection, and class imbalance problemsthat are inherent to intrusion detection. It also produces a new reliable dataset labeled Game Theory and Cyber Security (GTCS) that matches real-world criteria, contains normal and different classes of attacks, and reflects the current network traffic trends. The GTCS dataset is used to evaluate the performance of the different approaches, and a detailed experimental evaluation to summarize the effectiveness of each approach is presented. Finally, the thesis proposes an ensemble classifier model composed of multiple classifiers with different learning paradigms to address the issue of detection accuracy and false alarm rate in intrusion detection systems.
Dissertation or thesis originally submitted to ProQuest
Mahfouz, Ahmed Mosbah Elsaeed, "TOWARDS A HOLISTIC EFFICIENT STACKING ENSEMBLE INTRUSION DETECTION SYSTEM USING NEWLY GENERATED HETEROGENEOUS DATASETS" (2021). Electronic Theses and Dissertations. 2929.