PHISHING ATTACK PROGNOSIS USING MACHINE LEARNING

Phishing remains one of the most pervasive cyberthreats, exploiting human trust and subtle technical cues to steal credentials and sensitive data. This paper presents a machine-learning-driven prognosis framework designed to detect and predict phishing attacks earlier and with higher reliability than traditional rule-based systems. We combine multi-modal features drawn from URL and domain characteristics, email metadata and content (lexical and semantic embeddings), sender reputation, and lightweight user-behavior and network telemetry. After automated feature selection and engineering, several supervised models (Random Forest, XG Boost, Light GBM) and a sequence model (Bi-LSTM) are trained and ensembled to balance precision and recall. Experiments on multiple public and internal datasets show the proposed approach achieves strong detection performance (AUC > 0.95, F1 > 0.90) while reducing false positives compared to baseline heuristics. We also demonstrate a lightweight, low-latency inference pipeline suitable for deployment at mail gateways and endpoint agents, enabling near-real-time alerts and risk scoring for suspicious messages. Finally, we analyze feature importance to surface interpretable indicators that aid threat hunting and user education. The study shows that combining diverse signals with modern ML techniques produces an effective, explainable prognosis capability that can materially improve organizational resilience to phishing campaigns.