Performance Assessment of some Phishing predictive models based on Minimal Feature corpus

Abstract

Phishing is currently one of the severest cybersecurity challenges facing the emerging online community. With damages running into millions of dollars in financial and brand losses, the sad tale of phishing activities continues unabated. This led to an arms race between the con artists and online security community which demand a constant investigation to win the cyberwar. In this paper, a new approach to phishing is investigated based on the concept of minimal feature set on some selected remarkable machine learning algorithms. The goal of this is to select and determine the most efficient machine learning methodology without undue high computational requirement usually occasioned by non-minimal feature corpus. Using the frequency analysis approach, a 13-dimensional feature set consisting of 85% URL-based feature category and 15% non-URL-based feature category was generated. This is because the URL-based features are observed to be more regularly exploited by phishers in most zero-day attacks. The proposed minimal feature set is then trained on a number of classifiers consisting of Random Tree, Decision Tree, Artificial Neural Network, Support Vector Machine and Naïve Bayes. Using 10 fold-cross validation, the approach was experimented and evaluated with a dataset consisting of 10000 phishing instances. The results indicate that Random Tree outperforms other classifiers with significant accuracy of 96.1% and a Receiver’s Operating Curve (ROC) value of 98.7%. Thus, the approach provides the performance metrics of various state of art machine learning approaches popular with phishing detection which can stimulate further deeper research work in the evaluation of other ML techniques with the minimal feature set approach.