A hybrid approach on phishing URL Detection using Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) ”

Abstract

Phishing is one of the oldest types of cyber-attack which mostly comes in the form of camouflaged URLs to delude the users in order to get their personal information for malevolent purposes of the attacker. In addition, it is one of the easiest ways of inducing people into disclosing their personal credentials including credit card details. Since people use web applications on a daily basis, most phishing attacks comes up as fake websites pretending to mimic a trustworthy website. Moreover, emails are being used by the attackers to send the phishing website URL (Uniform Resource Locator) to the victim. Such type of URLs is termed as malicious URLs and most phishing attackers use them for successful data breaches. Therefore, it is a necessity to filter up, which URLs are benign, and which are malicious. In order to determine these factors, the concepts including traditional mechanisms used for URL detection, the drawbacks that those mechanisms had, and machine learning approaches used by different authors and their novelty approaches for effective detection are reviewed through this paper. Moreover, this will be focusing on cumulative deep learning approaches to build up hybrid deep learning models. Furthermore, this study proposes 4 hybrid deep learning models namely GRU-LSTM, LSTM-LSTM, bidirectional (GRU)-LSTM, and bidirectional (LSTM)-LSTM. In addition, the study also proposes 3 non hybrid deep learning models namely CNN(1D), LSTM and GRU. Hence, the main objective of this research is to provide a new insight to the hybrid deep learning approaches in URL detection by evaluating their accuracy, precision, recall and f1 score. In conclusion, this research recognizes Bi (GRU) – LSTM as the best mechanism to join hybrid models to detect phishing URLs and classify them as malicious or benign.