American
International
University-
Bangladesh

Students & Supervisors

Abstract

Phishing attacks remain a major cybersecurity concern. These attacks continue to evolve in complexity and frequently bypass traditional detection systems by imitating legitimate communication payloads. Many existing models particularly classical machine learning approaches struggle to detect hidden or adversarial phishing payloads and provide limited transparency in their predictions. This research introduces a phishing payload detection method using a fine-tuned DistilBERT model. The methodology includes dataset preprocessing, model fine-tuning, adversarial training, explainability analysis, and performance evaluation. DistilBERT, a lightweight transformer model, is fine-tuned to detect phishing payloads with enhanced accuracy and robustness. Adversarial training is integrated to defend against input manipulation, while explainable AI (XAI) techniques such as LIME and SHAP are employed to interpret the model’s predictions. The results show that DistilBERT achieves a classification accuracy of 98.52% and an AUC score of 0.9993, outperforming traditional machine learning models. It also maintains low false-positive rates and high recall. This research improves the reliability of phishing detection and provides interpretable outputs for security analysis. The findings demonstrate that the proposed framework strengthens phishing detection strategies and increases resilience to adversarial attacks. However, the results are based on a single publicly available phishing email dataset, and further validation across diverse datasets and real-world environments is required, with the scope of the findings limited to email-based phishing detection.

Keywords

Publication Details

• Type of Publication:
• Conference Name: IEEE Region 10 Conference 2025 (TENCON 2025)
• Date of Conference: 27/10/2025 - 27/10/2025
• Venue: Sabah International Convention Centre (SICC), Kota Kinabalu, Sabah, MALAYSIA
• Organizer: IEEE Malaysia Section