Evaluating the Efficacy of Machine Learning Methods in Phishing Detection: A Comparative Analysis

In the field of cybersecurity, the ongoing issue of phishing attacks is still of utmost concern because it puts both individuals and companies at serious risk. In order to prevent these fraudulent efforts and protect sensitive data, it is essential to implement effective detection and mitigation mechanisms. This research undertakes a thorough comparative evaluation of various machine learning models designed for phishing detection. Through a review of the literature on phishing detection techniques, we highlight the need for a methodical comparison analysis to identify the most effective strategies. In order to close this gap, we carefully assess the effectiveness of well-known machine learning methods, including Neural Networks, Gradient Boosting, XGBoost, Decision Trees, Random Forests, Support Vector Machines (SVM), and Logistic Regression. Using datasets from reliable sources, such as the PhishTank open-source service and the University of New Brunswick’s dataset, we create a balanced dataset with 5000 authentic URLs and the same amount of phishing URLs. Our approach includes thorough feature extraction and preprocessing steps to guarantee the accuracy and applicability of the dataset. To evaluate each model’s effectiveness, key performance indicators such F1-score, precision and recall are used. Our findings highlight subtle variations in the models' functionality and provide insight into their advantages and disadvantages for phishing detection. Additionally, we use visualisation approaches to explain each model’s outputs, which offer deeper insights into the operational dynamics of each model. Our results reveal subtle differences in the performance of the models, explaining their various advantages and disadvantages in the field of phishing detection. The resulting comparative study provides useful information about the effectiveness of machine learning techniques, which helps decision-makers in practical deployment settings.