Detecting AI-Generated Phishing Emails using Linguistic Patterns

     This study demonstrates that AI‑generated phishing emails bear distinct linguistic patterns that can be reliably detected. The combination of interpretable linguistic features and a fine‑tuned transformer yields a robust defence, with the feature‑based XGBoost model offering an attractive trade‑off between performance and explainability. Future work should extend the dataset to multilingual phishing, evaluate the impact of adversarial prompt engineering designed to mask linguistic signatures, and deploy the system as a plugin for enterprise mail gateways. As language models continue to evolve, the linguistic arms race between phishers and defenders will demand continuous retraining and novel detection paradigms, perhaps incorporating sender‑recipient relationship context alongside textual analysis.

  1. Verma, R., Shashidhar, N., Hossain, N.: Detecting Phishing Emails the Natural Language Way. In: Proc. European Symposium on Research in Computer Security (ESORICS), pp. 824–841 (2018) 2. Baki, S., Verma, R., Gnawali, O.: Scam Augmentation and Customization: Understanding the Linguistic Patterns of Phishing. In: Proc. IEEE Symposium on Security and Privacy Workshops, pp. 178–186 (2021) 3. Gehrmann, S., Strobelt, H., Rush, A.M.: GLTR: Statistical Detection and Visualization of Generated Text. In: Proc. ACL Demonstrations, pp. 111–116 (2019) 4. Guo, B., Zhang, Y., Xu, Q., et al.: HC3: A Suite of Test Collections for LLM-Generated Text Detection. In: Proc. ACL, pp. 12424–12441 (2023) 5. Tian, E.: GPTZero: An AI Text Detection Tool. https://gptzero.me (2023)