Adversarial Sample Detection with Gaussian Mixture Conditional Generative Adversarial Networks

<div>AUROC (%) of the threshold-based detector using the latent vector generated by semisupervised GAN’s generator at different basic blocks of ResNet trained on the CIFAR10 dataset and the logit vector of the target network. (a) FGSM. (b) BIM. (c) DeepFool. (d) CWL2. (e) PGD. (f) Feature ensemble.</div>

Mathematical Problems in Engineering

fig5

Figure 5

Figure 5: Adversarial Sample Detection with Gaussian Mixture Conditional Generative Adversarial Networks 