Adalward: a deep-learning framework for multi-class malicious webpage detection

Abstract

ABSTRACT A global village is what the father of digital media and communications, Marshall McLuhan had dreamt of in the late 1970s. In June 2017, reaching over 51.7% of the global population, the Internet has made it a reality. In past couple of decades, with upsurge of wireless communication technologies, the Internet has spread its web to connect all corners of the world. Its undeniable merits aside, interconnectivity on such a massive scale ushered in a whole new era of rampant malfeasance, characterized by an ever-increasing rate of cyber-crimes. Intrinsically, cyber-security researchers around the globe have been trying to develop several effective mechanisms to deal with the threats posed by cybercriminals. In this paper, we are presenting Adalward – a five-layer deep-learning framework, which has the potential for overcoming most of the challenges faced by such existing systems. Unique framework of Adalward allows it to utilize both static and dynamic web features for making accurate classification decisions with unmatched efficiency. Adalward was trained on one million labelled URLs obtained from numerous trustworthy sources. By the end of its training phase, Adalward achieved an overall detection accuracy of 99.76%, with a negligible false-positive rate of 0.10% and a nominal false-negative rate of 0.14%.