Evaluation of supervised machine learning methods in detection of phishing threats

This paper describes the development of seven machine learning models using the publicly available ISCX-URL2016 dataset. The models were designed to per-form multiple classifications, and their performance was evaluated using J48, Random Forest, Random Tree, Lazy IBk, BayesNet, and Naive Bayes algorithms. The dataset underwent preprocessing, resulting in 77 attributes, and the number of attributes for each model was determined using the InfoGain method. The results indicate that malicious website URLs can be classified into five predefined classes based on their features with high accuracy. The Random Forest, J48, Random Tree, and Lazy IBk algorithms achieved the highest accuracy rates, ranging from 94.52% to 98.43%. The Random Forest algorithm was further evaluated using machine learning metrics such as sensitivity, specificity, precision, recall, and f-measure, which demonstrated its effectiveness.