PRIME: A Phishing Detection Framework With Quantitative and Fuzzy-Based Dual Validation

Phishing attacks continue to pose a significant cybersecurity threat, especially as social engineering (SE) tactics become more contextually embedded and difficult to detect. To address the limitations of traditional rule-based framework and AI-driven classifiers, we propose PRIME, a phishing email evaluation framework that leverages large language models (LLMs) to assess manipulative intent across six interpretable criteria. Three risk scoring strategies, namely equal weighting, semantically weighted scoring, and fuzzy logic-based classification, are applied to aggregate the criterion scores into multi-level risk assessments. Qualitative comparisons with established frameworks demonstrate PRIME’s broad coverage and conceptual soundness. Quantitative experiments validate its effectiveness, with the fuzzy-based method achieving perfect recall on a phishing-only dataset and consistent performance across multiple years. An ablation study, where each criterion is removed in turn, highlights the critical role of the Context and Content dimension in detecting both explicit and subtle SE cues. By separating LLMs interpretation from final decision-making, PRIME enhances transparency, robustness, and adaptability in phishing detection systems.