Victory at the Robust Malware Detection Challenge
Our team, consisting of Laurens Bliek, Azqa Nadeem, Christian Hammerschmidt, and Sicco Verwer won both tracks of the Robust Malware Detection Challenge hosted by the Adversarial Machine Learning Workshop and sponsored by MIT-IBM Watson AI lab at KDD 2019.
The challenge addresses the difficulties of machine learning in the presence of adversaries in the context of malware detection. In today's modern cyber warfare, adversaries evade malware detectors by actively changing their malicious code until it is classified as benign, creating so-called adversarial examples. The goal of the challenge was (i) to propose a method for creating adversarial examples (attack track), and (ii) to build a new algorithm that learns how to make correct classifications in the presence of adversarial examples (defend track).
Our defend-track submissions proved to be the most robust against both the organizer's holdout malware samples as well as all competitors' submissions of crafted adversarial examples. The attack-track submission was the most successful at tricking all competitors' defense models. Details of the competition and results are available at https://sites.google.com/view/advml/advml19-challenge