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

Table 3

FNMRs in percent for a fixed FMR of 0.1%.

Model	Demographic group
	All	Female	Male	Dark	Light	df	dm	lf	lm
af_casia	4.341	5.539	4.093	2.525	4.987	2.874	2.517	5.632	4.799
af_glint360k	0.601	1.492	0.416	0.214	0.738	0.014	0.219	1.544	0.505
af_ms1mv2	0.561	1.449	0.377	0.211	0.686	0.014	0.216	1.499	0.450
af_ms1mv3	0.656	1.570	0.467	0.231	0.808	0.141	0.233	1.620	0.572
af_mxnet	0.568	1.523	0.369	0.207	0.696	0.014	0.211	1.576	0.440
af_webface600k	0.527	1.366	0.353	0.213	0.638	0.007	0.218	1.413	0.414
Curricularface	1.101	2.504	0.810	0.250	1.404	0.066	0.255	2.590	1.060
ef_arc	0.625	1.667	0.409	0.219	0.769	0.014	0.224	1.725	0.492
ef_arcplus	0.753	1.923	0.511	0.233	0.938	0.022	0.237	1.990	0.633
ef_cos	0.626	1.600	0.425	0.216	0.772	0.014	0.221	1.656	0.516
ef_cosplus	0.774	1.953	0.530	0.244	0.963	0.037	0.249	2.021	0.656
Magface	0.389	1.077	0.247	0.212	0.452	0.014	0.216	1.114	0.260
COTS	0.359	0.971	0.232	0.205	0.415	0.007	0.210	1.004	0.242

The lowest error rates are marked in bold.