LAYER-AWARE EFFICIENT FEATURE SELECTION FOR SOFTWARE-DEFINED MOBILE NETWORKING INTRUSION DETECTION

Abstract

Software-Defined Mobile Networking (SDMN) enables flexible 5G edge networks but introduces security challenges. Edge devices like base stations and gateways have limited computational resources, requiring lightweight intrusion detection solutions. Current intrusion detection evaluations prioritize accuracy while ignoring computational efficiency. Additionally, using global feature importance for feature selection hides whether different SDN architectural layers require different feature sets for effective attack detection. This thesis proposes a layer-aware adaptive feature selection framework using the SDN-ATK dataset, which contains 1,012,036 samples with 65 features and 19 classes. For feature selection, attacks are grouped into three SDN layers: Control Plane, Data Plane, and Application Plane. Three feature selection strategies are compared: Full using all 65 features, Fixed Union selecting the top ten features from each layer resulting in 18 features and Adaptive Union selecting features based on 90 percent cumulative importance per layer resulting in 48 features.