Phishing Detection Portal
Research project exploring the limits of traditional phishing detection systems through adversarial email generation, ReAct prompting strategies and Phi-2 QLoRA fine-tuning.
Naive Bayes Precision
97.6%
Detection After ReAct
15%
Phi-2 Detection Rate
22%
Model Size
2.7B
Evolution of Detection Rate
Research Overview
The project investigates how large language models can generate increasingly convincing phishing emails while simultaneously evaluating the robustness of classical detection systems.
A Naive Bayes classifier was trained as a baseline detector before introducing an adaptive ReAct loop where generated emails were continuously evaluated and reinjected into future prompts.
The second stage focused on fine-tuning Phi-2 using QLoRA 4-bit quantization, allowing efficient training while preserving strong language generation capabilities.