SeFACED: Semantic-Based Forensic Analysis and Classification of E-Mail Data Using Deep Learning
Maryam Hina, Mohsin Ali, Abdul Rehman Javed, Fahad Ghabban, Liaqat Ali Khan, Zunera Jalil
Abstract
Artificial Intelligence (AI), in combination with the Internet of Things (IoT), called (AIoT), an emerging trend in industrial applications, is capable of intelligent decision-making with self-driven analytics. With its extensive usage in diverse scenarios, IoT devices generate bulk data contrived by attackers to disrupt normal operations and services. Hence, there is a need for proactive data analysis to prevent cyber-attacks and crimes. To investigate crimes involving Electronic Mail (e-mail), analysis of both the header and the email body is required since the semantics of communication helps to identify the source of potential evidence. With the continued growth of data shared via emails, investigators now face the daunting challenge of extracting the required semantic information from the bulks of emails, thereby causing a delay in the investigation process. This gives an edge to the criminal in erasing their footprints of malicious acts. The existing keyword-based search techniques and filtration often result in extraneous, short sequence emails, which skips meaningful information. To overcome the above limitation, we propose a novel efficient approach named <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SeFACED</i> that uses Long Short-Term Memory (LSTM) based Gated Recurrent Neural Network (GRU) for multiclass email classification. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SeFACED</i> not only works on short sequences but with long dependencies of 1000+ characters as well. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SeFACED</i> focuses on tuning LSTM based GRU parameters to attain the best performance and with assessment by comparing it with traditional machine learning, deep learning models, and state-of-the-art studies on the subject. Experimental results on self-extended benchmark datasets exhibit that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SeFACED</i> effectively outperforms existing methods while keeping the classification process robust and reliable.