Email Spam Detection Using Machine Learning

Unsolicited emails, including phishing and spam, incur annual costs of millions of dollars for organizations and individuals. Numerous models and strategies for the automatic detection of spam emails have been introduced and developed; nevertheless, none have demonstrated 100% predictive accuracy. Both machine learning and deep learning algorithms demonstrated greater success among all proposed models. Natural language processing (NLP) improved the accuracy of the models. This study presents the efficacy of word embedding in the classification of spam emails. The pre-trained transformer model BERT (Bidirectional Encoder Representations from Transformers) is fine-tuned to identify spam emails vs non-spam (HAM). BERT employs attention layers to incorporate the context of the text into its analysis. The results are juxtaposed with a basic deep neural network model that incorporates a bidirectional Long Short Term Memory layer and two stacked dense layers. Furthermore, the results are juxtaposed with a collection of traditional classifiers, namely k-NN (k-nearest neighbors) and NB (Naive Bayes). Two open-source datasets are utilized, one for model training and the other for evaluating the model's persistence and resilience against novel data. The proposed method achieved a maximum accuracy of 98.67% and an F1 score of 98.66%.