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

Table 5

GARBE fairness evaluations for different demographics and different α factors, where  = 1 emphasizes on FNMR fairness and  = 0 emphasizes on FMR fairness.

Model	Gender			Skin color			Subgroups
	af_casia	0.150	0.284	0.419	0.327	0.520	0.713	0.237	0.479	0.720
af_glint360k	0.563	0.533	0.503	0.549	0.613	0.677	0.711	0.603	0.495
af_ms1mv2	0.586	0.414	0.241	0.528	0.507	0.487	0.716	0.643	0.571
af_ms1mv3	0.540	0.277	0.014	0.555	0.452	0.349	0.620	0.611	0.602
af_mxnet	0.609	0.469	0.329	0.541	0.467	0.349	0.730	0.555	0.380
af_webface600k	0.588	0.365	0.142	0.498	0.333	0.168	0.716	0.590	0.464
Curricularface	0.510	0.362	0.214	0.696	0.622	0.547	0.702	0.728	0.755
ef_arc	0.605	0.547	0.489	0.555	0.630	0.705	0.732	0.679	0.625
ef_arcplus	0.580	0.580	0.581	0.602	0.688	0.773	0.728	0.700	0.672
ef_cos	0.580	0.476	0.373	0.561	0.580	0.599	0.722	0.660	0.598
ef_cosplus	0.573	0.576	0.579	0.595	0.643	0.691	0.715	0.616	0.518
Magface	0.626	0.648	0.669	0.361	0.389	0.418	0.693	0.641	0.588
COTS	0.614	0.332	0.050	0.336	0.311	0.285	0.688	0.625	0.562

The best fairness values are marked in bold.