Hybrid AI Models for Detecting and Preventing Phishing Email

Abstract

As cybercriminals adopt advanced AI to launch "PhishBots"—automated tools that create highly realistic and personalized scam emails—traditional "black-box" security measures are failing to maintain user trust (Kumarage et al., 2025; Roy et al., 2023; Uddin & Sarker, 2024). This paper introduces a transparent, hybrid framework that combines the deep-learning power of RoBERTa-base with the SHAP interpretability engine. Our model achieves a peak accuracy of 99.43% on the PhreshPhish benchmark (Dalton et al., 2025; Meléndez et al., 2024).
By optimizing input sequences to 128 tokens, we achieve sub-100ms inference times, making it suitable for real-time enterprise deployment (Ferrag et al., 2023; Shirazi et al., 2022). This approach provides a "Digital Highlighter" for security analysts, transforming automated alerts into verifiable forensic evidence (Al‐Fayoumi et al., 2024; Lim et al., 2025).