Framework for detecting phishing crimes on Twitter using selective features and machine learning

Abstract

Socially aware information technology (SIT) plays a preferential role in facilitating the users for different tasks. Social media phishing is an escalating cybersecurity threat, where attackers employ deceptive tricks to steal personal data. Phishing detection in real-time is crucial, and highly dependent upon the selection of the most relevant features. Exiting literature often depends upon manual or random feature selection leading to inefficiencies in classification results. This research introduces a hybrid machine learning approach to phishing detection based on three feature selection methods Relief, Chi-square, and extra tree classifier for determining the most important features. Five classifiers including Naïve Bayes, support vector machine, decision tree, random forest (RF), and logistic regression are assessed based on accuracy, precision, recall, F1 score, and area under the curve (AUC). Experimental results indicate that RF obtains the highest accuracy of 95.56% and an AUC of 99.00%, better than other models and previous works. The results demonstrate the efficiency of the proposed method in improving phishing detection on social media.