Hybrid email filtering using TF-IDF and BERT enhanced spam detector

Abstract

Email spam detection is a crucial cybersecurity challenge requiring accurate and efficient filtering methods. This study introduces the TF-IDF and BERT Enhanced Spam Detector (TBESD), a hybrid deep learning model that integrates statistical and contextual text analysis for robust spam classification. TF-IDF captures essential n-gram patterns, while BERT embeddings extract deep semantic features. These feature vectors are combined and processed through a multilayer perceptron (MLP) with dropout regularization and ReLU activation for optimal classification. The methodology includes text preprocessing (normalization, stopword removal, tokenization), feature extraction (TF-IDF, BERT embeddings), and model training using the Adam optimizer with binary cross-entropy loss. Evaluated on a publicly available dataset, TBESD achieved an impressive accuracy of 98.88%, precision of 98.44%, recall of 99.44%, F1-score of 98.94%, and an AUC score of 0.9990. This study demonstrates the superiority of hybrid models over conventional methods, highlighting the role of advanced NLP techniques in cybersecurity.