The paper introduces Mal-AAEGAN, a generative framework designed to create adversarial malware samples. It builds on an adversarial autoencoder architecture and incorporates a Wasserstein GAN with gradient penalty (WGAN-GP) along with two critics. This places the work in a line of GAN-based malware generators that includes MalGAN, Mal-LSGAN, and the authors' earlier Mal-D2GAN.
The motivation is that machine-learning malware detectors are known to be vulnerable to adversarial samples. By generating such samples, researchers can probe these detectors and potentially identify ways to make them more robust. The double-critic setup and WGAN-GP are intended to stabilize training and improve the quality of the generated malware.
The abstract is truncated, so details on experimental results are not available in the source. The contribution appears to be a novel architecture for adversarial malware generation, announced as a new arXiv preprint.