افراح القيسي

Abstract In this thesis, we address security challenges specific to Direct-to-Cell connectivity in Low Earth Orbit satellite networks that cannot be resolved by conventional terrestrial network security solutions. We develop an intrusion detection system for LEO satellite networks using network emulation an‎d machine learning techniques. A realistic multi-satellite network topology was created an‎d validated against operational Starlink measurements from the LENS dataset to ensure emulated parameters accurately reflect real satellite network characteristics. The emulation generated a comprehensive dataset of 3,814,885 labelled network flows representing benign traffic an‎d nine attack types. Statistical analysis revealed less than six percent difference in key discriminative features between the simulated dataset an‎d the known 5G traffic dataset (5G-NIDD). Seven machine learning algorithms were eva‎luated after removing potential data leakage features including attack tool identifiers an‎d network addresses. Ensemble methods including Ran‎dom Forest, XGBoost, an‎d Decision Tree achieved near-perfect classification accuracy. LightGBM demonstrated optimal computational efficiency with training time of 12.45 seconds an‎d prediction time of 0.62 seconds for over one million test samples, establishing feasibility for deployment on resource-constrained satellite edge processors. Results demonstrate that intrusion detection systems trained on validated simulated data can achieve detection performance comparable to systems trained on real network traffic.

ماهواره‌هاي مدار پايين زمين , اتصال مستقيم به سلول , سيستم‌هاي تشخيص نفوذ , يادگيري ماشين , شبيه‌سازي شبكه , امنيت سايبري

Low Earth Orbit satellites , Direct-to-Cell connectivity , Intrusion detection systems , Machine Learning , network emulation , cyber security

Efrah Al Gheisi

Dr. Akbari