A hybrid ensemble deep learning framework with novel metaheuristic optimization for scalable malicious website...

Area of Science:

• Cybersecurity
• Machine Learning
• Data Science

Background:

• Malicious websites are a growing threat to online security.
• Traditional detection methods struggle to keep pace with evolving cyberattacks.

Purpose of the Study:

• To develop a hybrid detection framework for accurate malicious website identification.
• To integrate ensemble learning models with deep neural networks for enhanced detection capabilities.

Main Methods:

• Utilized a large dataset of 63,191 URLs with application-layer and network-layer features.
• Employed ensemble learning (Random Forest, XGBoost, LightGBM) and Deep Neural Networks.
• Applied Principal Component Analysis for dimensionality reduction and SHapley Additive exPlanations for model interpretability.
• Optimized hyperparameters using the Weevil Damage Optimization Algorithm and Energy Valley Optimizer.
• Validated model robustness with k-fold cross-validation.

Main Results:

• The hybrid framework significantly outperformed individual classifiers in accuracy and scalability.
• Achieved high accuracy in distinguishing malicious from benign websites.
• Demonstrated strong generalization capability and model interpretability.

Conclusions:

• The developed framework provides a reliable, data-driven, and explainable solution for real-time malicious website detection.
• Contributes to proactive cybersecurity by offering an advanced defense mechanism against evolving web threats.