Email Phishing Detection with BLSTM and Word Embeddings

Phishing has been one of the most successful attacks in recent years. Criminals are motivated by increasing financial gain and constantly improving their email phishing methods. A key goal, therefore, is to develop effective detection methods to cope with huge volumes of email data. In this paper, a solution using BLSTM neural network and FastText word embeddings has been proposed. The solution uses preprocessing techniques like stop-word removal, tokenization, and padding. Two datasets were used in three experiments: balanced and imbalanced, whereas in the imbalanced dataset, the effect of maximum token size was investigated. Evaluation of the model indicated the best metrics: 99.12% accuracy, 98.43% precision, 99.49% recall, and 98.96% f1-score on the imbalanced dataset. It was compared to an existing solution that uses the DL model and word embeddings. Finally, the model and solution architecture were implemented as a browser plug-in.