Revisiting Role Of Autoencoders In Adversarial Settings

Abstract

To combat against adversarial attacks, autoencoder structure is widely usedto perform denoising which is regarded as gradient masking. In this paper, werevisit the role of autoencoders in adversarial settings. Through thecomprehensive experimental results and analysis, this paper presents theinherent property of adversarial robustness in the autoencoders. We also foundthat autoencoders may use robust features that cause inherent adversarialrobustness. We believe that our discovery of the adversarial robustness of theautoencoders can provide clues to the future research and applications foradversarial defense.