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

Table 7

Fusion results based on GARBE candidates.

(a) Fusions of af_ms1mv3, af_mxnet, and af_webface600k to improve skin color fairness
	FMR	FNMR	FMR	FMR	GARBE
	All	All	Dark	Light	()
af_ms1mv3	0.1	0.656	0.267	0.129	0.349
af_mxnet	0.1	0.568	0.278	0.121	0.393
af_webface600k	0.1	0.527	0.215	0.153	0.168
AND-fusion	0.011	0.750	0.040	0.014	0.477
OR-fusion	0.242	0.441	0.572	0.328	0.271
Majority-Vote	0.047	0.562	0.149	0.062	0.416
Score-fusion	0.1	0.486	0.296	0.134	0.377
(b) Fusions of af_ms1mv3, af_webface600k, and curricularface 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
Curricularface	0.1	1.101	0.123	0.191	0.214
AND-fusion	0.005	1.197	0.007	0.016	0.430
OR-fusion	0.258	0.447	0.407	0.401	0.007
Majority-Vote	0.037	0.641	0.058	0.080	0.159
Score-fusion	0.1	0.510	0.145	0.230	0.227
(c) Fusions of af_glint360k, af_mxnet, and af_webface600k to improve fairness of demographic subgroups
	FMR	FNMR	FMR	FMR	FMR	FMR	GARBE
	All	All	Df	Dm	Lf	Lm	( = 1)
af_glint360k	0.1	0.601	0.748	0.531	0.108	0.155	0.495
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
AND-fusion	0.011	0.698	0.161	0.055	0.018	0.020	0.608
OR-fusion	0.242	0.445	1.470	0.826	0.355	0.560	0.375
Majority-Vote	0.046	0.554	0.487	0.214	0.070	0.091	0.532
Score-fusion	0.1	0.487	0.888	0.404	0.156	0.210	0.479

Bold numbers mark best results.