EURASIP Journal on Information Security

Table 2 Success probability and average L₂ distortion \({\overline {D}}\) when attacking networks adversarially trained against FGSM

From: Smooth adversarial examples

	MNIST - C4		ImageNet - InceptionV3
	P_suc	\({\overline {D}}\)	P_suc	\({\overline {D}}\)
FGSM	0.15	4.53	0.06	6.40
I-FGSM	1.00	3.48	0.97	29.94
PGD₂	1.00	2.52	1.00	3.89
C&W	0.93	3.03	0.95	6.43
qPGD₂	0.99	2.94	0.69	7.86
sC&W	0.99	2.39	0.75	6.22

Back to article page