FNet: A Two-Stream Model for Detecting Adversarial Attacks against 5G-Based Deep Learning Services

Table 6

Performance of normal images and their adversarial examples generated by FGSM on CIFAR-10.

Model		Method	Normal images		Adv images
			Precision	Recall	Precision	Recall
White model	VGG16	RGB-Net	0.896	0.928	0.864	0.807
		SRM-Net	0.748	0.773	0.571	0.538
		KDBU [32]	0.902	0.643	0.580	0.876
		FNet	0.926	0.926	0.868	0.868
Black model	ResNet	RGB-Net	0.888	0.928	0.874	0.809
		SRM-Net	0.692	0.773	0.543	0.440
		KDBU [32]	0.648	0.643	0.426	0.431
		FNet	0.912	0.926	0.879	0.854
	LeNet	RGB-Net	0.919	0.928	0.821	0.801
		SRM-Net	0.731	0.773	0.359	0.309
		KDBU [32]	0.697	0.643	0.269	0.319
		FNet	0.927	0.926	0.819	0.822

The bold values represent the results of experiments conducted by our method (FNet).