Research Article

On the Potential of Algorithm Fusion for Demographic Bias Mitigation in Face Recognition

Table 4

FMRs in percent when using the global system threshold on demographic subsets.

(a) Comparisons only within the same demographic group
ModelDemographic group
AllFemaleMaleDarkLightdfdmlflm

af_casia0.10.1170.2870.5210.0871.3530.5620.1220.017
af_glint360k0.10.1010.3070.4500.0860.7480.5310.1080.155
af_ms1mv20.10.1390.2270.3370.1161.1940.3730.1470.212
af_ms1mv30.10.1620.1670.2670.1290.8900.2960.1740.069
af_mxnet0.10.1210.2400.2780.1210.6960.3240.1300.401
af_webface600k0.10.1840.1380.2150.1530.6730.2390.2030.114
Curricularface0.10.1230.1910.3190.0931.6300.3150.1150.080
ef_arc0.10.1030.3000.4710.0811.1610.5380.1050.111
ef_arcplus0.10.0860.3270.5190.0661.3600.5780.0830.108
ef_cos0.10.1160.2550.3860.0961.1770.4210.1180.183
ef_cosplus0.10.0840.3150.4360.0790.9080.4960.0840.254
Magface0.10.2460.0480.0800.1950.5510.0860.2780.047
COTS0.10.1880.1700.2690.1501.0250.3060.2070.122

(b) Comparisons only across demographic groups
ModelAllGenderSkindf–dmdf–lfdf–lmdm–lfdm–lmlf–lm

af_casia0.10.0240.0110.3210.0370.0070.0090.0050.009
af_glint360k0.10.0300.0290.1560.0300.0260.0220.0640.026
af_ms1mv20.10.0280.0250.1540.0450.0360.0160.0520.032
af_ms1mv30.10.0300.0290.1250.0570.0250.0250.0310.025
af_mxnet0.10.0370.0350.0920.0350.0480.0200.1040.068
af_webface600k0.10.0230.0160.0980.0430.0190.0120.0200.037
curricularface0.10.0510.0540.2340.1230.0890.0410.0610.041
ef_arc0.10.0300.0290.1940.0500.0370.0220.0490.021
ef_arcplus0.10.0340.0340.2590.0570.0470.0230.0640.019
ef_cos0.10.0350.0340.2090.0630.0600.0220.0670.034
ef_cosplus0.10.0400.0400.2000.0470.0730.0240.1070.048
Magface0.10.0060.0050.0230.0200.0010.0030.0020.012
COTS0.10.0090.0060.0900.0300.0020.0020.0120.013

The lowest error rates are marked in bold.