Abstract:
The rapid evolution of cyber threats has forced the development of modern intrusion
detection systems (IDS) capable of identifying and combating sophisticated attacks.
Traditional IDS techniques often fail to adapt to changing threats, resulting in high
false-positive rates and insufficient accuracy. This study presents a robust intrusion
detection framework employing ensemble learning techniques, specifically stacking, to
boost cybersecurity threat detection. The research leverages the UNSW-NB15 dataset,
a baseline for testing IDS, comprising multiple attack types and normal network traffic.
The stacking ensemble combines Random Forest, Gradient Boosting, and XGBoost as
foundation models with Logistic Regression as a meta-learner, providing a model that
capitalizes on the complimentary qualities of its components. Rigorous preprocessing
and feature engineering techniques are utilized to refine the dataset and increase model
performance. Evaluation criteria, including accuracy, precision, recall, F1-Score,
indicate the superiority of the suggested model. The stacking ensemble obtains an
accuracy of 96.7%, a precision of 95.8%, and an F1-Score of 95.6%, greatly surpassing
single models. The False Positive Rate is decreased to 2.1%, illustrating the model’s
practical effectiveness in lowering false alarms and assuring reliable threat detection.
This research emphasizes the potential of ensemble learning in boosting the
adaptability, scalability, and resilience of IDS, addressing major concerns in modern
cybersecurity. The findings provide a platform for establishing sophisticated, real-time
detection systems and pave the way for future breakthroughs in intrusion detection
approaches.
