Adversarial Sample Detection with Gaussian Mixture Conditional Generative Adversarial Networks

Table 2

Partially supervised scenarios in detecting adversarial samples.

Model	Dataset	Method	Partially unsupervised scenario
			FGSM	BIM	DeepFool	CW	PGD
DenseNet	CIFAR10	Mahalanobis	99.94	99.51	83.42	87.95	81.84
		Ours (latent)	99.95	90.79	98.06	95.75	76.00
		Ours (score)	99.98	92.23	96.09	97.18	78.00
	CIFAR100	Mahalanobis	99.86	98.27	75.63	86.20	39.32
		Ours (latent)	100.00	89.86	83.14	79.08	62.35
		Ours (score)	99.89	90.15	80.19	81.09	64.15
	SVHN	Mahalanobis	99.85	99.12	93.47	96.95	81.40
		Ours (latent)	99.95	99.00	98.71	98.16	94.15
		Ours (score)	99.90	99.01	98.29	97.18	92.16
ResNet	CIFAR10	Mahalanobis	99.94	98.91	78.06	93.90	100.00
		Ours (latent)	99.98	73.19	96.79	95.71	100.00
		Ours (score)	99.87	96.15	94.13	94.10	100.00
	CIFAR100	Mahalanobis	99.77	96.38	81.95	90.96	99.85
		Ours (latent)	99.92	81.18	83.32	86.63	100.00
		Ours (score)	99.93	80.10	80.13	87.01	100.00
	SVHN	Mahalanobis	99.62	95.39	72.20	86.73	99.92
		Ours (latent)	99.96	74.89	95.97	89.65	99.96
		Ours (score)	99.65	75.14	95.10	89.75	99.23