Research on Multimodality Face Antispoofing Model Based on Adversarial Attacks

Table 1

The accuracy of single-modality and multimodality models under attacks.

Data stream type	Original sample (%)	FGSM, 0.01 perturbation (%)	DeepFool (%)	NATTACK (%)
RGB	94.1	2.5	8	24.3
Depth	98.5	63.9	4.9	7.6
IR	96.3	1	10.7	12.4
Fusion	99.5	2.7	0.7	31.5