Saudi Cultural Missions Theses & Dissertations
Permanent URI for this communityhttps://drepo.sdl.edu.sa/handle/20.500.14154/10
Browse
1 results
Search Results
Item Restricted A MULTI-LAYER DEFENSE FRAMEWORK FOR ENHANCING ADVERSARIAL ROBUSTNESS OF MACHINE LEARNING-BASED INTRUSION DETECTION SYSTEMS(Saudi Digital Library, 2026) alshmmri, Eiman Salem; Almatarneh, Rami JibreelThe rapid evolution and increasing sophistication of cyber threats, especially the adversary attacks have revealed significant weaknesses in the traditional Intrusion Detection Systems (IDS). The traditional signature-based systems can no longer identify a zero-day attack or respond to an ever-changing pattern of threats. Even though the Intrusion Detection Systems based on Machine Learning (ML-IDS) have enhanced detection performance, most of the existing systems are not robust in the adversarial settings and are commonly tested at fixed and static settings. This restricts the knowledge of model resilience, stability and generalization in real and unstable cybersecurity conditions. This thesis will solve those issues by introducing a single, configuration-sensitive assessment system of ML-based intrusion detectors. The framework combines tree models, such as Random Forest, Extra Trees, and XGBoost, and an optimized ensemble method to improve predictive performance, stability, and generalization. It also includes feature selection as a way of minimizing redundancy and enhancing efficiency and class imbalance management as a way of stabilizing learning when the data distribution is skewed. Moreover, the adversarial perturbation testing is done both in white-box and black-box to represent real-life attack conditions. A multi-level defense mechanism is also implemented to enhance model resilience and reduce the effect of adversarial attack. The framework is tested on two non-homogeneous benchmark datasets, namely the CIC-IDS2018, published in 2018, and the updated release of the ToN-IoT dataset, update published in 2025. It allows assessment in various time frames and changing threat patterns, so that the offered approach will be highly efficient in both well-established and recently re-evaluated data sets. The experimental results demonstrate that the proposed framework achieves a high level of performance, with accuracy exceeding 95% in many cases, alongside strong F1- score and AUC values across different attack categories. The ensemble model exhibits enhanced stability and generalization capability, while feature selection and class imbalance handling contribute significantly to improving efficiency and robustness. Moreover, the proposed defense mechanism effectively reduces performance degradation in adversarial environments and enhances system recovery. These findings highlight the effectiveness of the proposed framework in developing robust, adaptive, and resilient Intrusion Detection Systems capable of maintaining high performance under diverse and evolving cybersecurity conditions.5 0