Email Spam Detection: A Novel Hybrid Approach Using Machine and Deep Learning Techniques

Email spam continues to evolve as a persistent cybersecurity threat, employing sophisticated obfuscation techniques that challenge conventional detection systems.This paper introduces a fundamentally new approach to spam classification through three key innovations: (1) a dual-path feature extraction system that uniquely combines CNN-based spatial analysis of text with DFT-based frequency-domain pattern recognition, enabling detection of both lexical and obfuscation patterns; (2) a dynamic feature fusion mechanism that optimally weights textual and spectral features based on email structure; and (3) an adaptive thresholding classifier using XGBoost that automatically adjusts decision boundaries for imbalanced data.Unlike prior hybrid models that simply concatenate features, our framework implements hierarchical attention to prioritize high-risk indicators while suppressing noise -a critical advancement for handling modern adversarial spam.The proposed system demonstrates state-of-the-art performance across three benchmark datasets (Ling-Spam, Enron, SpamAssassin), achieving 99% accuracy and 1.00 precision on Ling-Spam while maintaining 98% accuracy on larger corpora.Crucially, it reduces false negatives on obfuscated content by 28% compared to BERT-based models and improves cross-dataset generalization by 15% over conventional hybrids, as validated through rigorous statistical testing (Wilcoxon p<0.01).The DFT integration proves particularly effective against character substitution attacks, with spectral analysis identifying 92% of obfuscated spam that escapes lexical detection.