Fraud E-mail detection using intelligent algorithms: Comparison of traditional approaches with deep learning techniques

Abstract

Fraud emails pose a persistent cybersecurity threat by tricking recipients into disclosing sensitive information. This study evaluates and compares the performance of traditional machine learning and deep learning techniques for fraud email detection using a publicly available dataset containing 17,538 emails. Features were extracted using Term Frequency-Inverse Document Frequency (TF-IDF). Traditional models including Naive Bayes, Logistic Regression, XGBoost, and Random Forest achieved up to 98.52 % accuracy, while deep learning models like Bi-LSTM and GRU reached a maximum accuracy of 97.71 %. Evaluation metrics such as confusion matrices, ROC curves, and AUC scores were used for comprehensive performance comparison. Results demonstrate that traditional models can outperform deep learning models on text-based email data with proper feature engineering, offering efficient and scalable solutions for fraud detection systems.