Replacing Human Input in Spam Email Detection Using Deep Learning

Abstract

The Covid-19 pandemic has been a driving force for a substantial increase in online activity and transactions across the globe. As a consequence, cyber-attacks, particularly those leveraging email as the preferred attack vector, have also increased exponentially since Q1 2020. Despite this, email remains a popular communication tool. Previously, in an effort to reduce the amount of spam entering a users inbox, many email providers started to incorporate spam filters into their products. However, many commercial spam filters rely on a human to train the filter, leaving a margin of risk if sufficient training has not occurred. In addition, knowing this, hackers employ more targeted and nuanced obfuscation methods to bypass in-built spam filters. In response to this continued problem, there is a growing body of research on the use of machine learning techniques for spam filtering. In many cases, detection results have shown great promise, but often still rely on human input to classify training datasets. In this study, we explore specifically the use of deep learning as a method of reducing human input required for spam detection. First, we evaluate the efficacy of popular spam detection methods/tools/techniques (freeware). Next, we narrow down machine learning techniques to select the appropriate method for our dataset. This was then compared with the accuracy of freeware spam detection tools to present our results. Our results showed that our deep learning model, based on simple word embedding and global max pooling (SWEM-max) had higher accuracy (98.41%) than both Thunderbird (95%) and Mailwasher (92%) which are based on Bayesian spam filtering. Finally, we postulate whether this improvement is enough to accept the removal of human input in spam email detection.