Enhancing Phishing Detection Using O3-Mini Chain-of-Thought Reasoning with GPT-4o

Choudary, Bharath Reddy (2025) Enhancing Phishing Detection Using O3-Mini Chain-of-Thought Reasoning with GPT-4o. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (640kB) | Preview
Preview	PDF (Configuration Manual) Download (553kB) | Preview

Abstract

The phishing attacks are still highly sophisticated using advanced social engineering techniques that outplay traditional detection methods. While deep learning approaches shown some hope, they act more as black boxes, lacking interpretability required for security applications. This paper discusses a new way of phishing detection enhancement through enriched GPT-4o and chain-of-thought reasoning by OpenAI's o3-mini model.

The experimental design is a three-stage one based on 10,000 balanced emails from the corpus of Enron. In the first phase, o3-mini generates structured argumentation chains assessing five aspects of security, i.e., sender legitimacy, linguistic characteristics, employment of methods of social engineering, technical characteristics, and risk assessment. The argumentation chains are also used for fine-tuning of GPT-4o, thus having a model that makes classification judgments simultaneously and also generates well-explained descriptions. The testing pits four setups against each other: baseline GPT4o, few-shot GPT-4o, reasoning-augmented fine-tuned GPT-4o, and o3-mini in isolation.

The result indicates that the method enhanced by rationality achieves 99.0% accuracy, an increment by 18.4 percentage points from baseline GPT-4o (80.6%), while exhibiting a 94.5% overall error minimization. Precision and recall are balanced equally at 99.0% by the fine-tuned model, and no classification bias is observed in comparison to baseline configurations.

The proposed method achieves 99.0% accuracy compared to the baseline's 80.6%, reducing errors by 94.5% and maintaining a consistent score of 99.0% for both recall and precision.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Garg, Mohit UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing H Social Sciences > HV Social pathology. Social and public welfare > Criminology > Crimes and Offences > Cyber Crime Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	28 May 2026 13:35
Last Modified:	28 May 2026 13:35
URI:	https://norma.ncirl.ie/id/eprint/9317

Actions (login required)

View Item