A Generative Adversarial Network Based Approach to Malware Generation Based on Behavioural Graphs

Abstract

As the field of malware detection continues to grow, a shift in focus is occurring from feature vectors and other common, but easily obfuscated elements to a semantics based approach. This is due to the emergence of more complex malware families that use obfuscation techniques to evade detection. Whilst many different methods for developing adversarial examples have been presented against older, non semantics based approaches to malware detection, currently only few seek to generate adversarial examples for the testing of these new semantics based approaches. The model defined in this paper is a step towards such a generator, building on the work of the successful Malware Generative Adversarial Network (MalGAN) to incorporate behavioural graphs in order to build adversarial examples which obfuscate at the semantics level. This work provides initial results showing the viability of the Graph based MalGAN and provides preliminary steps regarding instantiating the model.