Answers are grounded in this paper and cite their sources.
Comparative analysis of recurrent neural networks reveals BiLSTM detects 91.43% of phishing emails, suggesting its superiority in email security tasks.
Purpose of the study: to consider the features of the use of recurrent neural networks in solving the problem of detecting phishing emails. Methods of research: comparison, mathematical and software modeling, system analysis. Result(s): The concept and types of phishing attacks are considered. The analysis of modern publications on the use of recurrent neural networks in phishing detection tasks has been carried out, which has shown that the use of recurrent networks makes it possible to detect phishing emails with high probability. Publicly available datasets have been analyzed: most datasets are focused on detecting phishing URLs. The few datasets focused on the text of an email are overwhelmingly in English, and high-quality Russian-language datasets are not publicly available, so our own dataset of Russian-language emails was compiled. Mathematical and software modeling of various recurrent neural networks for detecting phishing emails has also been carried out: RNN, LSTM, BiLSTM and a comparative analysis of their characteristics has been carried out. The dependences of loss characteristics and accuracy on the number of epochs are revealed. A comparative analysis of recurrent networks has shown that the BiLSTM network, which detected 91.43 % of phishing emails, was the most effective in solving phishing detection problems in the framework of research. The RNN network showed the worst characteristics, which detected only 50.71 % of phishing emails from the test sample. It should be noted that these results were obtained for networks trained on small-volume datasets (300 emails). Scientific novelty: the research results allow us to reasonably conclude that of the considered recurrent neural networks, BiLSTM is the one that best copes with the tasks of detecting phishing emails with small amounts of training dataset.
No takes yet. Share an insight, caveat, or question.
Boldyrikhin et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: