Phishing Detection Using Ensemble of Classifiers

Phishing remains to be a major cybersecurity threat despite its long history. We explore URL-based phishing detection using ensemble methods and recursive feature selection. URL-based approaches extract features from different substrings of the URL. It is also common for approaches to use external features such as indexing and time domain response. Our results suggest that for ensembles, the parameters substring does not contribute to phishing detection while the directory substring consistently provides important features. Google-based features are also not important to ensembles. Our best performing ensemble is an AdaBoost classifier which uses both external and substring features, with an accuracy of 93.74% and 0.00001 s average prediction time.