Abstract
The rapid advances in machine learning and deep learning algorithms have led to their adoption to tackle different security problems such as spam, intrusion, and malware detection. Malware is a type of software developed with a malicious intent to damage, exploit, or disable devices, systems, or networks. Malware authors typically operate through black-box sitting when they have a partial knowledge about the targeted detection system. It has been shown that supervised machine learning models are vulnerable to well-crafted adversarial examples. The application domain of malware classification introduces additional constraints in the adversarial sample crafting process compared to the computer vision domain: (1) the input is binary and (2) retaining the visual appearance of the malware application and its intended functionality. In this paper, we have developed a heterogeneous ensemble classifier that combines supervised and unsupervised models to hinder black-box attacks designed by two variants of generative adversarial network (GAN). We experimentally validate its soundness on a corpus of malware and legitimate files.