Adversarial Sample Detection with Gaussian Mixture Conditional Generative Adversarial Networks

Table 3

Unsupervised scenarios for detecting adversarial samples.

Model	Dataset	Method	FGSM	Unsupervised scenario
				BIM	DeepFool	CW	PGD
DenseNet	CIFAR10	Odds-testing	45.23	69.01	58.30	61.29	97.93
		GM-CGAN	87.89	73.69	80.81	78.12	46.45
	CIFAR100	Odds-testing	43.22	65.22	49.53	47.64	96.91
		GM-CGAN	98.09	68.34	74.42	65.72	41.09
	SVHN	Odds-testing	56.14	71.11	67.81	70.71	99.25
		GM-CGAN	83.35	73.56	81.86	80.16	46.40
ResNet	CIFAR10	Odds-testing	46.32	59.85	75.58	57.58	96.18
		GM-CGAN	96.68	64.74	79.48	73.49	98.80
	CIFAR100	Odds-testing	38.26	43.52	61.13	44.74	93.73
		GM-CGAN	80.96	81.20	80.35	67.56	91.30
	SVHN	Odds-testing	65.09	70.31	77.05	72.12	99.08
		GM-CGAN	94.97	89.04	94.71	83.41	97.52