الفهرس | Only 14 pages are availabe for public view |
Abstract In today’s digitally-transformed world, the Internet and information networks play a vital role in virtually every aspect of people’s lives, and in various sectors, including education, healthcare, finance, and defence. However, this digital revolution has also brought about significant challenges, particularly concerning the security and integrity of these networks. With the increasing reliance on Internet-based services and the exponential growth of data volumes, malicious actors have emerged, seeking to exploit vulnerabilities and manipulate sensitive information for personal gain or malicious intent. The consequences of successful cyber-attacks can be devastating, leading to financial loss, privacy breaches, disruption of critical services, and even threats to national security. To ensure the continuous availability, confidentiality, and integrity of data in this interconnected landscape, safeguarding the networks against cyber threats has become an absolute necessity. Among the essential components of network security, Intrusion Detection Systems (IDSs) play a pivotal role in detecting and mitigating potential threats. IDSs monitor network traffic in real-time, analysing patterns and behaviours to identify anomalies that may indicate unauthorized access attempts, malicious activities, or suspicious behaviour. By promptly alerting network administrators to potential threats. IDSs enable timely response and mitigation measures, minimizing the impact of cyber-attacks and ensuring the continuous operation of critical systems. However, the everevolving nature of cyber threats necessitates the development of intelligent and adaptive defence mechanisms. Traditional IDSs often struggle to keep pace with the rapidly changing landscape of attack techniques, making it imperative for network security researchers to explore innovative approaches that can enhance the effectiveness of IDSs. This thesis presents three automated anomaly detection models using Convolutional Neural Networks (CNNs) to augment the capabilities of IDSs. The first model is trained from scratch using the NSL-KDD intrusion detection dataset, while the other two models are built based on pre-trained models, namely the Visual Geometry group method using 19 layers (VGG19) and the Residual Network using 152 layers (ResNet152), with the UNSW-NB15 database serving as the evaluation dataset. By leveraging the power of Deep Learning (DL) methodology, the proposed models demonstrate remarkable accuracy in detecting network intrusions concurrently with minimizing the false predictions. Through extensive experimentation and comprehensive comparisons with state-of-the-art IDS models, the effectiveness and superiority of the proposed models are substantiated. Ultimately, this research contributes to the ongoing efforts of securing Internet and information networks in the age of digital transformation. |