Adaptive Phishing URL Detection Using Feature Stability and Uncertainty-Based Decision Making

hishing URLs are designed to mislead users into visiting deceptive websites and disclosing sensitive information. Blacklistbased systems can identify previously reported addresses, but newly created or modified URLS may not yet have a reputationrecord. Machine-learning classifiers can identify patterns in URL strings; however, their performance may depend on datasetcomposition, feature selection, and changes in attacker behaviour. This paper proposes an adaptive phishing URL detection pipelinethat combines lexical and structural feature extraction, feature-stability analysis, supervised classification, and uncertainty-baseddecision making. Feature stability is assessed across repeated or time-separated training samples to identify features that remainconsistently informative. The classifier produces a Phishing or Legitimate label only when its confidence satisfies a validation-selectedthreshold; otherwise, the system returns Uncertain for further review. The paper defines an evaluation protocol using precision,recall, F1-score, ROC-AUC, calibration, coverage, and covered error rate, with domain- and time-aware splits where feasible. Thiswork presents a proposed methodology and research design. The model has not yet been implemented or experimentally evaluated,and no quantitative performance improvement is claimed.