OPTIMIZING SVM FOR PHISHING DETECTION: A COMPARATIVE PERFORMANCE STUDY

Abstract

Phishing attacks continue to pose significant cybersecurity threats, particularly through deceptive emails designed to steal user credentials or deliver malware. Existing detection systems rely heavily on machine learning models such as Support Vector Machines (SVM), whose performance is sensitive to hyperparameter optimization. This study investigates how different optimization techniques improve the performance of an SVM-based phishing email detection model. Four approaches were compared using a labelled phishing Kaggle dataset: a baseline linear SVM, a Grid Search–optimized SVM, a Stochastic Gradient Descent (SGD)–based SVM, and a Bayesian-optimized SVM with an RBF kernel. Each model was evaluated using standard performance metrics. The baseline SVM achieved 92.87% accuracy; Grid Search reached 96.00%; SGD achieved 92.00%; and Bayesian Optimization achieved the highest accuracy of 96.67%. Hyperparameter optimization significantly enhances SVM performance in phishing detection, with Bayesian Optimization offering the most efficient and accurate configuration.