Research Article
On the Potential of Algorithm Fusion for Demographic Bias Mitigation in Face Recognition
Table 6
Fusion results based on FMR candidates.
| (a) Fusions of ef_arcplus and Magface to improve skin color fairness | | FMR | FNMR | FMR | FMR | GARBE | | | | All | All | Dark | Light | () | | |
| ef_arcplus | 0.1 | 0.753 | 0.519 | 0.066 | 0.773 | | | Magface | 0.1 | 0.389 | 0.080 | 0.195 | 0.418 | | | AND-fusion | 0.009 | 0.782 | 0.030 | 0.013 | 0.398 | | | OR-fusion | 0.191 | 0.361 | 0.570 | 0.249 | 0.392 | | | Score-fusion | 0.1 | 0.393 | 0.275 | 0.147 | 0.302 | | |
| (b) Fusions of ef_cosplus and Magface to improve gender fairness | | FMR | FNMR | FMR | FMR | GARBE | | | | All | All | Female | Male | () | | |
| ef_cosplus | 0.1 | 0.774 | 0.084 | 0.315 | 0.579 | | | Magface | 0.1 | 0.389 | 0.246 | 0.048 | 0.669 | | | AND-fusion | 0.007 | 0.804 | 0.014 | 0.014 | 0.006 | | | OR-fusion | 0.193 | 0.360 | 0.317 | 0.351 | 0.051 | | | Score-fusion | 0.1 | 0.395 | 0.194 | 0.158 | 0.102 | | |
| (c) Fusions of af_casia, ef_rcplus, and Magface to improve fairness of demographic subgroups | | FMR | FNMR | FMR | FMR | FMR | FMR | GARBE | | All | All | df | dm | lf | lm | (ā=ā1) |
| af_casia | 0.1 | 4.341 | 1.353 | 0.562 | 0.122 | 0.017 | 0.720 | ef_arcplus | 0.1 | 0.753 | 1.360 | 0.578 | 0.083 | 0.108 | 0.672 | Magface | 0.1 | 0.389 | 0.551 | 0.086 | 0.278 | 0.047 | 0.588 | AND-fusion | 0.002 | 4.431 | 0.068 | 0.010 | 0.004 | 0.001 | 0.842 | OR-fusion | 0.277 | 0.349 | 2.758 | 1.114 | 0.447 | 0.166 | 0.628 | Majority-Vote | 0.021 | 0.704 | 0.441 | 0.104 | 0.034 | 0.008 | 0.778 | Score-fusion | 0.1 | 0.408 | 1.400 | 0.357 | 0.201 | 0.044 | 0.702 |
|
|
Bold numbers mark best results.
|