Phishing Emails: Analysis and Detection with Comparison of Three Machine Learning Models (LR, NB and MLP)

Phishing attacks are performed by writing and forwarding falsified body of email messages which look legitimate or real from an undisputed origin to a victim or different category of victims. They focus at acquiring the sensitive data of users or by transferring and loading malware on the user's computers. Consequently, this study aims to implement an AI-driven approach to detect emails that seems to be phishing while the features are analyzed. This project leverage three machine learning models namely: MLP, a deep learning algorithm, Naïve Bayes and Logistic Regression. The following performance metrics were obtained -> for Multilayer Perceptron (MLP) model: accuracy: 98.57%, precision: 100.00%, recall: 90.00% while the f1_score metrics was 94.74%. For Naïve Bayes (NB) model: accuracy: 96.95%, precision: 100.00%, recall: 78.75% while the f1_score metrics was 88.11%. For Logistic Regression (LR) model: accuracy: 94.71%, precision: 99.03%, recall: 63.75% while the f1_score metrics was 77.57%. The result shows that MLP Classifier may better capture complex patterns in phishing emails, leading to higher detection rates. Naive Bayes is still a strong choice, especially for simpler or smaller datasets due to its speed and efficiency. Logistic Regression is reliable but slightly less accurate on this particular task. For this project, a phishing email dataset from the Kaggle Machine Learning Repository was utilized. This dataset contains 5000+ instances of phishing and ham emails.