Intelligent Models for Mitigating the Impact of Phishing on Electronic Commerce Data: A Critical Narrative Review of Detection Capability, Robustness and Operational Evidence

Abstract

Phishing remains the dominant entry point through which the personal, financial and transactional data held by electronic commerce platforms are compromised, and the research response has been overwhelmingly computational. Several hundred studies now propose intelligent models based on feature-engineered classifiers, deep neural architectures, multimodal reference-based systems and, most recently, large language models. Reported classification performance is consistently high, yet losses attributable to online shopping fraud and credential theft have not declined correspondingly. This review examines that discrepancy. Its purpose is to evaluate critically what the available evidence establishes about the ability of intelligent models to reduce data loss in electronic commerce settings, as distinct from their ability to separate labelled samples within curated datasets. Literature was identified through open scholarly databases and citation searching, appraised for methodological adequacy, and synthesised thematically around five problems: the mapping between detection outputs and the data assets actually at risk; the comparative evidence for competing model families; the construction and temporal validity of evaluation datasets; robustness under adversarial and distributional pressure; and the conditions under which a detection decision becomes a mitigation. The evidence indicates that headline performance figures are strongly conditioned by dataset construction, that accuracy on balanced benchmarks translates poorly to the extreme base-rate asymmetry of live traffic, and that robustness has been assessed for only a small proportion of published models. Reference-based and multimodal designs show more stable behaviour under impersonation than purely lexical models, but at a computational cost that is rarely reported in terms compatible with transaction-time constraints. Language-model detectors improve semantic sensitivity and explanation quality while introducing latency, cost and privacy trade-offs that remain sparsely characterised. Evidence linking model deployment to reduced data compromise is almost entirely absent. Priorities include temporally partitioned and platform-realistic evaluation, standardised adversarial reporting, and outcome measures defined in terms of data exposure rather than classification alone.