Identifying and Evaluating the Best Machine Learning Predictive Models for Detecting Voice (Phone-Call) Vishing Attacks on MoMo Users in Real Time

Abstract

Phishing, particularly voice-based phishing (vishing), has become a significant security threat, exploiting human trust and the widespread use of mobile communication. This paper aims to develop and evaluate a hybrid model that combines Gradient Boosting and Convolutional Neural Networks (CNNs) for detecting phishing calls in audio data. The hybrid model leverages the strengths of Gradient Boosting, a powerful classification technique, and CNNs, which excel at extracting features from raw audio signals. To assess the model’s effectiveness, a dataset comprising 40 phishing audio files and 40 legitimate audio files was used. The audio data was converted into spectrograms for CNN training. Experimental results indicate that the hybrid approach outperforms individual models such as Gradient Boosting and CNNs, assessing performance based on precision, recall, accuracy, and ROC AUC. The model specifically achieved an accuracy of 70.83%, with 67% precision for phishing calls and 75% precision for legitimate calls. By combining traditional machine learning with deep learning, this study presents an innovative approach to phishing detection. The findings highlight the effectiveness of integrating advanced feature extraction methods with robust classification techniques to enhance security in mobile money platforms. The proposed hybrid model offers a promising solution for real-time vishing detection, with potential applications in securing financial transactions and improving fraud prevention mechanisms.