Email Spam Detection: An Adaptive, Multi-Modal Ensemble Framework Utilizing Machine Learning and Zero-Shot Heuristics

Combatting spam and phishing emails continues to be among the most difficult challenges in cybersecurity. Traditional filtering approaches are incapable of coping with the increasing sophistication of attackers' techniques. The current research proposes an advanced and adaptable AI-powered framework which combines five proven machine learning algorithms — Multinomial Naïve Bayes, SVM, Random Forest, KNN, and Multi-Layer Perceptron — along with specialized heuristics for identifying zero-day phishing attacks and image spam. The unique feature of the framework is a dynamic weighting scheme combined with continuous learning capability, wherein the system retrains itself upon user-reported misclassifications. The architecture utilizes semantic understanding based on zero-shot learning principles applied to large language models. The proposed approach achieves over 98% accuracy on a balanced multi-modal dataset.