Adversarial Example Attacks and Defenses in DNS Data Exfiltration

The Domain Network System (DNS) protocol is used on a daily basis to access the internet. It acts as a phone book that allows users to access websites using words rather than remembering address numbers. In recent years it has become clear that there are serious vulnerabilities in the DNS protocol, and the lack of attention to these vulnerabilities (e.g. data exfiltration) is concerning. The widespread use of the DNS protocol opens a door that could possibly allow for companies and users to have their data stolen through data exfiltration. Machine learning is a popular tool for malicious traffic detection, however they are vulnerable to adversarial examples. This leads to the security arms race, where researchers aim to accurately detect and counter new malicious threats exploited by malicious actors. In this work, we demonstrate the success of adversarial examples of DNS exfiltration packets in bypassing machine learning detection techniques. We then propose a voting ensemble method to improve adversarial attack detection. The voting ensemble proposed increases the accuracy of adversarial detection, providing a new level of protection against adversarial example attacks.