Lightweight and Scalable Real-Time Phishing URL Detection Models: A Comparative Review and Proposed Framework

Abstract: Phishing URL detection has advanced considerably through machine learning and deep learning, with recent published systems reporting accuracy figures that frequently exceed 95%. However, a close reading of this literature shows that high accuracy is often achieved using large feature sets, deep sequence architectures, or ensembles comprising a dozen or more classifiers — approaches that are difficult to justify on resource-constrained platforms such as browser extensions, mobile applications, or edge gateways, where real-time response and a small memory footprint matter as much as raw detection accuracy. This paper reviews six recent studies on machine-learning-based phishing URL detection, comparing their algorithms, feature counts, and reported accuracy, and identifies a consistent gap: none of the reviewed works measure deployment cost — inference latency, model size, or throughput under concurrent load — alongside their accuracy claims. Building on this analysis, the paper proposes a tentative methodology for a lightweight, scalable, real-time phishing-detection framework that restricts features to those computable from the URL string alone, benchmarks compact classical classifiers against the heavier baselines identified in the review, and treats latency, model size, and throughput as first-class evaluation metrics rather than afterthoughts. The proposed direction is intended to narrow the gap between laboratory-grade accuracy and field-deployable, real-time phishing protection.