The Role of Feature Selection in Machine Learning for Detection of Spam and Phishing Attacks

Abstract

With the increase in Internet use throughout the world, expansion in network security is indispensable since it decreases the chances of privacy spoofing, identity or information theft and bank frauds. Two of the most frequent network security breaches involve phishing and spam emails as they are an easy way to pass a virus or a malicious site, which can lead to extensive frauds. Despite the fact that there is an abundance of tools for detection and blocking of these types of messages and websites, society is still trying to combat and rise above said problem. The purpose of this paper was to exclude the human factor in security breaches executed in this manner with the use of various machine learning algorithms. For the purpose of training and testing of the most successful algorithms (Random Forest, k-Nearest Neighbor, Artificial Neural Network, Support Vector Machine, Logistic Regression, Naive Bayes) paper used two separate bases, UCIs Phishing Websites Data Set and Spam Emails Dataset together with Weka software, and found that the best results for both of them are achieved with the Random Forest algorithm. However, databases responded differently to feature selection algorithms, as the best result for phishing (97.33% accuracy) was accomplished through Ranker + Principal Components Optimization, and the best result for spam (94.24% accuracy) was accomplished through BestFirst + CfsSubsEval Optimization in Weka. These findings provide a base platform for future work towards a faster and more accurate online fraud detection.