Worst-Case Morphs Using Wasserstein ALI and Improved MIPGAN
Table 3
MMPMR for FRLL morphs.
Landmark-based morphing
GAN-based morphing
Worst- case
AMSL
Face- Morpher
OpenCV
WebMorph
Style- GAN
MIPGAN
Improved MIPGAN (Ours)
WALI (Ours) MFN&EF w/o colourcorr.
WALI (Ours) MFN&EF w colourcorr.
MobileFaceNet
64.7
89.2
86.0
90.3
22.1
74.4
96.2
96.7
96.8
99.5
ElasticFace
38.8
58.8
60.4
60.0
0.0
4.7
74.4
26.1
44.0
99.8
Curricularface
32.7
53.4
55.6
54.3
0.0
2.6
44.3
3.0
11.0
99.6
ArcFace
58.8
67.2
64.4
65.3
1.2
20.3
64.6
7.8
17.0
99.4
Inception
9.6
14.2
14.2
17.3
0.3
5.0
11.2
1.2
2.4
49.3
PocketNet
51.7
78.0
80.3
84.7
20.0
55.9
75.8
14.1
44.5
98.8
VGG16
12.1
12.8
14.2
25.2
0.4
6.8
20.5
2.3
4.7
100.0
Dlib
0.9
1.0
1.0
0.9
0.0
0.0
0.6
0.1
0.1
16.5
COTS
97.6
100.0
100.0
100.0
0.8
70.4
96.6
46.6
64.7
n/a
Note. The more challenging the morphs, the higher the MMPMR. Underlined numbers indicate the FR system was used for optimisation. Of all GAN-based approaches the improved MIPGAN approach (highlighted in bold) is closest to the worst case for almost all FR systems.