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

Table 8

Fusion results based on Pareto-efficient candidates.

(a) Fusions of af_webface600k and Magface to improve skin color fairness
	FMR	FNMR	FMR	FMR	GARBE
	All	All	Dark	Light	()
af_webface600k	0.1	0.527	0.215	0.153	0.168
Magface	0.1	0.389	0.080	0.195	0.418
AND-fusion	0.017	0.573	0.026	0.030	0.076
OR-fusion	0.183	0.344	0.270	0.319	0.084
Score-fusion	0.1	0.383	0.141	0.182	0.127
(b) Fusions of af_ms1mv3, af_webface600k, and Magface to improve gender fairness
	FMR	FNMR	FMR	FMR	GARBE
	All	All	Female	Male	()
af_ms1mv3	0.1	0.656	0.162	0.167	0.014
af_webface600k	0.1	0.527	0.184	0.138	0.142
Magface	0.1	0.389	0.246	0.048	0.669
AND-fusion	0.006	0.737	0.014	0.009	0.217
OR-fusion	0.257	0.323	0.507	0.292	0.270
Majority-Vote	0.036	0.514	0.073	0.054	0.145
Score-fusion	0.1	0.388	0.213	0.119	0.280
(c) Fusions of af_mxnet, af_webface600k, and Magface to improve fairness of demographic subgroups
	FMR	FNMR	FMR	FMR	FMR	FMR	GARBE
	All	All	df	dm	lf	lm	( = 1)
af_mxnet	0.1	0.568	0.696	0.324	0.130	0.401	0.380
af_webface600k	0.1	0.527	0.673	0.239	0.203	0.114	0.464
Magface	0.1	0.389	0.551	0.086	0.278	0.047	0.588
AND-fusion	0.006	0.656	0.110	0.018	0.014	0.005	0.724
OR-fusion	0.259	0.317	1.424	0.535	0.523	0.508	0.308
Majority-Vote	0.035	0.512	0.387	0.098	0.075	0.051	0.562
Score-fusion	0.1	0.391	0.922	0.239	0.229	0.111	0.541

Bold numbers mark best results.