Worst-Case Morphs Using Wasserstein ALI and Improved MIPGAN

Table 4

Detection performance in BPCER (%) at APCER 5% and 10% (Section 5.1).

Trained with	Land-; mark	MIPGAN	Improved MIPGAN; (Ours)	WALI (Ours) with optimisation using:							WALI (Ours) 512  512; baseline
				MFN	MFN &; El.Face	MFN &; Curr.Face	MFN &; ArcFace	MFN &; Inception	MFN &; PocketNet	MFN &; VGG16
S-MAD	BPCER@APCER 5%
FRGC	3.2	100.0	99.4	71.5	66.2	67.0	69.2	70.9	71.2	63.9	81.3
AMSL	38.7	100.0	95.2	54.2	62.4	48.8	45.9	48.4	48.0	42.6	99.8
	BPCER@APCER 10%
FRGC	1.4	100.0	98.6	56.0	52.2	50.7	55.9	49.8	45.5	47.9	78.8
AMSL	26.2	100.0	89.0	29.5	38.1	31.4	26.5	26.9	26.7	26.9	99.4
D-MAD	BPCER@APCER 5%
FRGC	0.3	19.4	18.5	7.2	9.8	10.2	12.8	8.3	6.7	8.4	0.5
	BPCER@APCER 10%
FRGC	0.2	12.4	12.0	3.5	5.1	5.6	7.3	4.8	3.7	4.2	0.2

Note. Top = LBP-based S-MAD. Bottom = D-MAD based on FR-difference features.