Enhancing Cybersecurity through Stacked Ensemble Learning Approach for Multiclass URL Classification

In today's digital era, detecting malicious URLs is essential to avoid substantial financial losses from inadvertent clicks on harmful links. This research presents a sophisticated method employing stacked ensemble learning for multiclass URL classification. Through Grid Search with Cross-Validation, hyperparameter tuning is performed on Bernoulli Naive Bayes and Multinomial Naive Bayes models. These optimized models act as base estimators in a Stacking Classifier, with Logistic Regression serving as the meta-classifier, achieving an impressive 98% accuracy. The technique leverages a substantial dataset of 651,191 URLs, categorized as benign, defacement, malware, and phishing, using TF-IDF vectorization and Random Over-sampling to address class imbalance. Evaluation metrics such as accuracy, classification report, and confusion matrix demonstrate the model's efficiency in strengthening cybersecurity measures.