Defense against adversarial attacks using DRAGAN

Ali ArjomandBigdeli, Maryam Amirmazlaghani, Mohammad Khalooei

Abstract: In recent years, deep neural networks have been used in a wide range of applications such as machine vision, speech recognition, natural language processing, etc., and have achieved significant success, however, these networks are vulnerable to adversarial attacks. This has raised concerns about the security of these networks. In this paper, we are going to use Generative Adversarial Networks (GANs) to resist image classifiers against a range of adversarial attacks. To do this, we took the inspiration of Defense-GAN and improving it by using DRAGAN (Deep Regret Analytic Generative Adversarial Networks). We train the GAN on unperturbed images, and after that, the GAN will be used to reconstruct the input images before feeding to the image classifier, which makes it resistant to adversarial attacks.