AI-Enabled Phishing Detection: A Comparative Study of Learning-Based Approaches

Abstract

Phishing remains a major cybersecurity problem because deceptive websites, messages, and links can imitate legitimate services while changing rapidly. This paper presents an expanded comparative review of learning-based approaches for phishing detection. It examines conventional blacklists, whitelists, and rule-based mechanisms alongside machine learning, deep learning, natural-language processing, screenshot analysis, and hybrid approaches. Particular attention is given to URL structure, domain characteristics, webpage content, textual signals, visual similarity, feature extraction, dataset construction, and evaluation methodology. The discussion shows that learning-based methods can generalize beyond previously reported malicious addresses, but their practical effectiveness depends on representative and current datasets, meaningful features, balanced evaluation, and resilience to changing attack patterns. Deep and multimodal systems may provide richer representations, while lightweight models remain useful for low-latency browser and mobile deployment. The paper also discusses concept drift, adversarial manipulation, shortened URLs, external feature dependencies, explainability, privacy, and computational cost. Finally, a conceptual adaptive framework is proposed in which multiple evidence sources are combined and confirmed observations are used for controlled model updates.