Research on Multimodality Face Antispoofing Model Based on Adversarial Attacks

Table 6

The accuracy of the model before and after adding the defence module to the multimodality model.

	The original authors’ pretrained model (%)	After adding the diffusible high-frequency suppression (%)	After using the mixed adversarial training (%)	Combination of the high-frequency suppression module and the mixed adversarial training (%)
Original	99.5	97.58	96.8	99.7
FGSM 0.01	2.7	25.8	95.2	92
DeepFool	0.7	64.1	93.5	81.4
NATTACK	31.5	68	91.9	96