The problem
Phishing is the first entry point of most attacks on universities and companies. Rule-based filters miss new campaigns; large cloud models would see every private e-mail. A small fine-tuned language model on the organisation's own server can be accurate, fast and private.
Who needs it
University and company IT departments, e-mail security vendors.
Idea of the approach
- Public phishing and legitimate e-mail datasets (and, with permission, anonymised reported phishing from campus IT).
- Baselines: classic features + logistic regression (as in the early work), then fine-tune a small BERT-style model with LoRA.
- Highlight the words or links that triggered the alert.
- Measure detection rate, false alarms and time per e-mail on a CPU server.
What you will build and deliver
- A filter service with explanations and a test interface.
- Evaluation report and paper.
Skills needed
Python, NLP basics, Hugging Face; e-mail formats.
Research contribution & where to publish
Contribution: accuracy–speed–explanation trade-offs of small fine-tuned models for phishing, on CPU only. Venues: IEEE TrustCom, eCrime (APWG), Computers & Security.
Key references — start reading here
- I. Fette, N. Sadeh, A. Tomasic. Learning to detect phishing emails. WWW, 2007.
- J. Devlin, M.-W. Chang, K. Lee, K. Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL-HLT, 2019.
- E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen. LoRA: Low-Rank Adaptation of Large Language Models. ICLR, 2022.
Interested?
Send the topic title and the courses you have taken.
Contact Dr. Hamza Djigal