An ML-Based Solution to Detect and Classify Suspicious E-Mails

Abstract

The e-mail service is a powerful communication tool that is utilized in both private and corporate settings all over the world. However, for more than two decades, organizations and individuals have been subjected to spam/phishing e-mails, which are regarded as nuisances. The widespread of suspicious emails, also known as spam, has created a need for reliable and robotics antispam filters. Therefore, many spam filtering systems recently utilize Machine Learning-based approaches to detect spam emails effectively. For example, several antivirus programs, such as Cylance, Deep Instinct, Avast, and Windows Defender Security, have great AI and ML implementations. In this thesis, we intend to use different Machine Learning based techniques used in spam email detection to protect the e-mail inbox from receiving potentially harmful e-mails. Other than that, Explainable Artificial Intelligence (XAI) techniques were utilized to provide explanations for the decisions of black-box Machine Learning techniques. Both attackers and defenders could utilize the explanations provided by XAI techniques and this thesis would focus on how to mitigate and defend against spam attackers. More precisely, a testing dataset based on the XAI explanations will be created as XAI cyber attackers and the defensive framework will work as XAI cyber defenders. Attackers utilized word substitution for texts and masking blocks for images based on the information provided by XAI. And defenders would utilize a novel scaffolding technique that effectively hides the biases of any given classifier by allowing an adversarial entity to craft an arbitrary desired explanation.

Date of Award	Apr 2023
Original language	American English
Supervisor	ERNESTO Damiani (Supervisor)