Feature Selection for Improved Phishing Detection

Phishing – a hotbed of multibillion dollar underground economy – has become an important cybersecurity problem. The centralized blacklist approach used by most web browsers usually fails to detect zero-day attacks, leaving the ordinary users vulnerable to new phishing schemes; therefore, learning machine based approaches have been implemented for phishing detection. Many existing techniques in phishing website detection seem to include as many features as can be conceived, while identifying a relevant and representative subset of features to construct an accurate classifier remains an interesting issue in this particular application of machine learning. This paper evaluates correlation-based and wrapper-type feature selection techniques using real-world phishing data sets with 177 initial features. Experiments results show that applying an effective feature selection procedure generally results in statistically significant improvements in the classification accuracies of—among others—Naïve Bayes, Logistic Regression and Random Forests, in addition to improved efficiency in training time.