ARCHIVES

Phishing attacks represent one of the most persistent and damaging cybersecurity threats in the modern digital landscape, systematically exploiting human cognitive vulnerabilities to illicitly obtain sensitive information including login credentials, financial account data, and personal identity details. Conventional rule-based and blacklist-driven detection systems have demonstrated a pronounced inability to adapt to the rapidly evolving sophistication of contemporary phishing techniques, resulting in elevated false-positive rates, significant missed detections, and an ongoing reliance on labour-intensive manual maintenance. This paper presents a comprehensive AI-driven phishing detection framework that systematically integrates Natural Language Processing (NLP) and Machine Learning (ML) methodologies to substantially enhance both detection accuracy and operational robustness. The proposed system incorporates multi-stage text preprocessing, hybrid feature extraction combining Term Frequency–Inverse Document Frequency (TF-IDF) vectorisation and pre-trained word embeddings including Word2Vec and GloVe, alongside a comparative evaluation of supervised classification models encompassing Logistic Regression, Support Vector Machines (SVM), Random Forest, and Long Short-Term Memory (LSTM) deep learning networks. Experimental evaluation conducted across a combined dataset of 129,382 labelled email samples demonstrates that the proposed hybrid NLP-ML model substantially outperforms both traditional rule-based approaches and single-method ML baselines, with the LSTM classifier achieving 96.7% accuracy, 96.3% precision, 96.0% recall, and an F1-score of 96.1%. The principal contributions of this work include a rigorous comparative analysis of six machine learning architectures, a scalable and modular detection pipeline suitable for real-time deployment, a comprehensive feature importance analysis identifying key discriminative attributes, and actionable insights for enhancing operational phishing detection systems.