Leveraging stacking machine learning models and optimization for improved cyberattack detection

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Complex cyber attacks necessitate advanced intrusion detection systems (IDS).
• Existing methods struggle with high-dimensional data, imbalanced classes, and high false positive rates.
• Robust and accurate IDS are crucial for network defense.

Purpose of the Study:

• To introduce an innovative framework for intrusion detection.
• To address challenges in accuracy, imbalanced data, and efficiency in current IDS.
• To develop a superior IDS by integrating novel machine learning techniques.

Main Methods:

• Integration of J48 and ExtraTreeClassifier machine learning models for classification.
• An improved Equilibrium Optimizer (EEO) for feature selection using K-Nearest Neighbors (KNN) Fisher and accuracy scores.
• Synthetic Minority Oversampling Technique with Iterative Partitioning Filters (SMOTE-IPF) for class balancing and KNN for data imputation.

Main Results:

• Achieved high accuracy (99.7% on NSL-KDD, 98.1% on UNSW-NB15) and F1 scores (99.6% and 98.0% respectively).
• Demonstrated superior performance in feature selection precision and classification accuracy.
• Effectively handled minority class instances and showed improved computational efficiency.

Conclusions:

• The proposed IDS framework significantly enhances detection capabilities.
• The methodology offers a robust solution for managing imbalanced datasets and reducing false positives.
• The system provides a computationally efficient and accurate approach to network intrusion detection.