The Defense of Adversarial Example with Conditional Generative Adversarial Networks

Figure 3

Experimental results of the defense adversarial example in the MNIST dataset. The first line shows clean images. The second line shows an adversarial example. Here, we use four different approaches to generate adversarial (CW, DeepFool, FGSM, and JSMA) images and three images for each approach. The third line shows the images generated by the defense framework.