Efficient and Lightweight Phishing Detection: A Case for Sustainable Cybersecurity with TF-IDF and LightGBM

Phishing is a damaging cyber threat in the digital landscape, and AI-generated attacks are becoming increasingly sophisticated. This paper presents a study on the development and evaluation of a computationally lightweight and efficient model for phishing email detection. An approach using TF-IDF + LightGBM was proposed, achieving 95.8% recall and 90.4% precision on a real-world dataset of 18,650 emails, reducing training time by 98% and memory requirements by 99% compared to state-of-the-art deep learning models (BERT, DistilBERT). Our methodology covers the complete machine learning pipeline, from data acquisition and exploratory analysis to rigorous hyperparameter tuning and performance evaluation. We demonstrate that an optimized classical ML approach can achieve performance comparable to deep learning models while being orders of magnitude more efficient. This work addresses the need for robust, efficient, and interpretable defenses that can be deployed in resource-constrained environments. We further contribute an analysis of model failures, feature importance, and robustness against adversarial attacks. Our findings demonstrate the value of democratizing access to advanced cybersecurity technologies through efficient and sustainable AI solutions.