A Robust and Explainable Transformer-Based Framework for Phishing Email Detection

AI models need built-in defenses against adversarial attacks: The paper demonstrates that standard phishing detection models can be fooled by deliberately crafted perturbations, but adding adversarial training using techniques like the Fast Gradient Method significantly improves resilience against these sophisticated attacks.

Explainability is essential for trust in AI security systems: By integrating three different explanation methods (LIME, SHAP, and Integrated Gradients), the framework makes it possible for security teams to understand why an email was flagged as phishing, rather than treating the AI as a "black box"—critical for high-stakes security decisions.

Lightweight models can be both effective and transparent: The use of DistilBERT (a smaller version of BERT) shows that you don't need massive, resource-intensive AI models to detect phishing effectively; smaller models can be more interpretable while maintaining strong performance.

Plain-language explanations bridge the gap between AI and users: By using Flan-T5-Small to convert technical AI explanations into understandable language, the framework makes security insights accessible to non-technical stakeholders who need to act on phishing alerts.

Robustness and explainability work together: The paper shows that combining adversarial training (to make models more resilient) with explainability methods (to make decisions transparent) creates a more trustworthy system than either approach alone.