RPCP-PURI: A robust and precise computational predictor for Phishing Uniform Resource Identification

Abstract

The widespread growth of internet services has led to a substantial increase in traffic across multifarious online applications such as shopping, transportation, e-commerce, and banking. Cyber-criminals are stealing internet users credentials through Phishing Uniform Resource Locator (PURL). To make sure safe usage of web services, researchers have developed several AI-supported PURL predictors. Prime motivation behind development of each new predictor was to extract comprehensive features from URLS and more accurately discriminate benign and phishing URLs. Different predictors are enriched with diverse types of feature extraction strategies and deep potential exploration of these strategies remain unexplored. With an aim to develop more powerful PURL predictor, contributions of this manuscript are manifold: It brings different sets of feature extraction strategies at single platform and explores their individual as well as combine potential. It evaluates and compare the effectiveness of BERT language model for detection of phishing URLs (PURLs). It presents a novel LSTM and CNN-based hybrid predictor, which utilizes DeepWalk-generated URL embeddings to effectively discriminate between phishing and benign URLs. Extensive experimentation across three public benchmark datasets (EBBU2017, phishing sites URLs and Phishing and benign webpages) demonstrates that proposed hybrid predictor surpasses the performance of existing predictors by 0.2%, 1.9% and 1.2% respectively.