Hyperparameter of Catboost classifier.

Web applications are important for various online businesses and operations because of their platform stability and low operation cost. The increasing usage of Internet-of-Things (IoT) devices within a network has contributed to the rise of network intrusion issues due to malicious Uniform Resource Locators (URLs). Generally, malicious URLs are initiated to promote scams, attacks, and frauds which can lead to high-risk intrusion. Several methods have been developed to detect malicious URLs in previous works. There has been a good amount of work done to detect malicious URLs using various methods such as random forest, regression, LightGBM, and more as reported in the literature. However, most of the previous works focused on the binary classification of malicious URLs and are tested on limited URL datasets. Nevertheless, the detection of malicious URLs remains a challenging task that remains open to research. Hence, this work proposed a stacking-based ensemble classifier to perform multi-class classification of malicious URLs on larger URL datasets to justify the robustness of the proposed method. This study focuses on obtaining lexical features directly from the URL to identify malicious websites. Then, the proposed stacking-based ensemble classifier is developed by integrating Random Forest, XGBoost, LightGBM, and CatBoost. In addition, hyperparameter tuning was performed using the Randomized Search method to optimize the proposed classifier. The proposed stacking-based ensemble classifier aims to take advantage of the performance of each machine learning model and aggregate the output to improve prediction accuracy. The classification accuracies of the machine learning model when applied individually are 93.6%, 95.2%, 95.7% and 94.8% for random forest, XGBoost, LightGBM, and CatBoost respectively. The proposed stacking-based ensemble classifier has shown significant results in classifying four classes of malicious URLs (phishing, malware, defacement, and benign) with an average accuracy of 96.8% when benchmarked with previous works.