MAG-GAN: Massive attack generator via GAN

Abstract

Adversarial attacks reveal the vulnerability of deep neural networks (DNNs). These attacks fool DNNs by adding small perturbations to normal examples. Currently, most attacks involve generating a single example or target a single deep model. To gain insight into adversarial attacks and develop a robust defense, this study focuses on a generic attack model applicable to most adversarial attacks. A novel mass-generator of adversarial examples with a strong attack ability and involving small perturbations is presented herein. The main contributions of this work include proposing a generic framework for adversarial attacks, designing comprehensive evaluation metrics for adversarial examples, and developing a novel method for mass-generating adversarial examples via a generative adversarial network (MAG-GAN). Finally, experiments were conducted to demonstrate the good performance of MAG-GAN compared with state-of-the-art attack methods. Once the model was trained, adversarial examples were mass-generated with a small perturbation and a strong attack ability. Furthermore, it was found that MAG-GAN model can be adopted as an efficient tool to reveal the vulnerability and improve the defense ability of existing DNNs. A promising result is that the target model mounted in MAG-GAN exhibited a good defense performance after game training, which is equivalent to adversarial training.