Training curves with score-based and false matching rate-based objective functions. Figure 8a shows training curves for four master faces, two generated using different objective functions, guided only by feedback from the 2D FR system, and the other two guided only by the 3D FR system. As shown in Figure 8a, these two different objective functions achieved similar FMRs in the 2D-only scenario. For 3D FMR, the score-based function performed better. However, Figure 8b shows that the score-based function failed to jointly attack the 2D and 3D FR systems. After 1,000 iterations, the FMR-based function has an FMR of 6.6%, while the score-based function holds only 0.06%.