Razložljivo zaznavanje anomalij v omrežni varnosti: primerjalna študija modelov strojnega učenja in transformatorjev
Year: 2026
Type: dissertation
Abstract: Modern networks face attacks resembling benign traffic, making anomaly detection critical. Beyond accuracy, explainability is essential so security teams can trust model outputs. This study compares traditional machine learning and transformer-based tabular models for network-intrusion detection, investigating which deliver strong detection on minority attack classes while remaining interpretable. Using a customized dataset based on UNSW-NB15 (The University of New South Wales) with approximately 251,000 flows and light preprocessing addressing multi-class distribution, we evaluate Logistic Regression, Random Forest, Extreme Gradient Boosting, Categorical Boosting, a Multi-Layer Perceptron, and tabular transformers (TabTransformer, Self-Attention and Intersample Attention Transformer, Tabular Prior-Fitted Networks version 2, Tabular Kolmogorov-Arnold Network) using a single 80-20 stratified train-test split, with explanations from SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations. Traditional methods outperformed transformers in accuracy and balanced detection. SHAP analysis showed traditional models generate interpretable explanations while transformers produce values too small to be directly interpretable, requiring additional post-processing before they become useful to analysts, with cross-model consensus identifying Time To Live values and connection timing as key attack indicators. Practical deployment requires hybrid strategies: traditional models for real-time operations with immediate explanations, and transformers for specialized scenarios accepting computational overhead.
Source: RUNG
Author: Cheloei, Shayan
Institution: University of Nova Gorica
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164707440
