MSc dissertation comparing transformer NLP models (DistilBERT, BERT-base) against traditional ML baselines (Naïve Bayes, Logistic Regression, Random Forest) for phishing email detection in cloud environments. Positivist, quantitative, experimental design — evaluated using accuracy, F1-score, and false-negative rate as the security-critical metric. BERT-base achieved the best result: 0.34% FNR on 155,039 emails.
My MSc dissertation: 'Design and Implementation of an AI-Powered Threat Detection System for Cloud Communication Platforms'. The core question was how effectively different machine learning approaches — from classical models to modern transformers — can detect phishing emails in a large, real-world dataset, and what trade-offs each approach involves for production deployment.
BERT-base achieved the best result with an FNR of 0.34%, with DistilBERT close behind at 0.40%. Both transformer models substantially outperformed the classical baselines on FNR, at the cost of higher training time and inference latency — a trade-off discussed in the dissertation's evaluation chapter as a key consideration for production deployment.
Transformer models were deployed as HuggingFace checkpoints and classical models serialised as pickles, with a Gradio dashboard built for live demonstration of the detection pipeline.
The pipeline architecture here — ingestion → feature extraction → classification → alert — generalises directly to production use cases such as security monitoring, fraud detection, and anomaly detection systems, and is the direct continuation of the Smishing Framework capstone.