A Comparative Study Of Artificial Intelligence-Based Models To Curb Prevalence Of Cybercrime

Abstract

Background: Phishing remains a pervasive threat in the realm of cybersecurity, necessitating effective detection mechanisms to safeguard individuals and organizations from malicious attacks. However, deep learning can be used to effectively combat this evolving threat landscape following its success story in other fields of application.     

Objective: For this study therefore, we seek to investigate the potentials of one-dimensional CNN-based phishing detection (1D-CNNPD) model and other existing deep learning techniques, with a view to addressing known challenges against accuracy of phishing email detection.  

Methods: We carried out an experiment to assess the capability of the 1D-CNNPD models and developed an augmentation using a Long Short-Term Memory (LSTM) and a Gated Recurrent Unit (GRU) in order to give additional efficiency to its detection ability. Further experiment was carried out using two standard datasets as benchmark.                                                                             

Results: It was discovered that appreciable model enhancement was achieved with additional layers to the base 1D-CNNPD model. The two benchmark datasets include the Spam Assassin, and the Phishing Corpus datasets. Performance results indicate that the 1D-CNNPD with Bi-GRU augmentations outperforms DeepAnti-PhishNet as well as other similar deep learning models, which were found to achieve better phishing email detection accuracy.          

Conclusion: Our discovery that the Bi-GRU augmentation alone could achieve detection accuracy of 99.72% and an F1 score of 99.68% implies that the proposed enhancement is about 0.40% more efficient than the second-best performing model.