A Feature Guided Denoising Network For Adversarial Defense

As neural networks are playing a more and more significant role in many fields, they are also under the risk of being attacked by adversarial examples, which becomes a great challenge to the whole community. Due to this problem, networks cannot provide the guaranteed security when they are used in some security-sensitive scenarios. In this paper, a feature guided denoising network for adversarial defense is studied. Specifically, a denoising network is designed to remove possible adversarial perturbations on the input image. And a novel training method of the denoising network is proposed to improve the performance, in which deep features extracted from clean examples by the pretrained classifier is used as supervision information in the training process. Experimental results reveal that the proposed method shows satisfactory performance on defending against several white-box adversarial attacks. Besides, combination of the proposed method and adversarial training is studied, which achieves very good results compared to the other experiments reported in this paper.