Advances in Phishing Detection: A Comprehensive Review of Machine Learning, Deep Learning, and Transformer-based Approaches

Phishing remains one of the most pervasive and damaging forms of cyber-enabled fraud, driving credential theft, account takeover (ATO), ransomware ingress, and large-scale supply-chain compromise. Escalating attacker sophistication, ranging from phishing-as-a-service (PhaaS) kits to AI-generated spear phishing, has outpaced purely reactive defenses, such as blacklists and rule-based filters. In response, research has shifted toward automated phishing detection pipelines leveraging machine learning (ML), deep learning (DL), representation learning, and, more recently, Transformer/LLM-driven techniques capable of contextual and semantic reasoning. This review provides a comprehensive survey across the phishing-defense stack: attack taxonomies; data sources and benchmarking datasets; classic ML baselines; CNN/RNN advances; Transformer-era breakthroughs (BERT variants, URLTran-class models, lightweight/distilled architectures); explainability, adversarial robustness, and deployment constraints; ecosystem signals such as SPF/DKIM/DMARC and brand impersonation detection; and socio-technical considerations including user awareness and reporting loops. We synthesize results from peer-reviewed publications and industry threat reports and offer techniques for next-generation, adaptive, privacy-preserving, and multimodal phishing defense systems.