Abstract
Passport officers and Automated Border Control (ABC) gates are susceptible to face-morphing attacks, wherein a blended/morphed image of two different identities is mistaken for either original identity (Nightingale et al., 2021). Face identification is also affected by the Other-Race Effect (ORE), in which own-race faces are identified more accurately than other-race faces. We examined the ORE in face-morph identification. East Asian (EA) and Caucasian (CA) participants (N=60) completed a face-identity matching test for pairs of EA and CA face images (32 same-identity, 32 different-identity image pairs). In the morph condition, different-identity pairs consisted of a natural image of identity A with a 50% morph of identities A and B. Same-identity pairs consisted of an image of A and a morph of two different images of A. In the baseline condition, face pairs were made from unedited images. Participants indicated whether image pairs showed the same or different identities, using a 5-point scale (1: sure same; 5: sure different). Performance was measured using the area-under-the ROC curve (AUC). Performance was less accurate for morph pairs than baseline pairs (p< .001). A 3-way interaction between participant race, face-image race, and condition (p=0.038) showed that East Asians identified EA morphs more accurately (AUC=0.73) than CA morphs (AUC=0.70) and Caucasians identified CA morphs (AUC=0.77) more accurately than EA morphs (AUC=0.70). A deep convolutional neural network (DCNN) (Ranjan et al., 2018) trained for face identification attained an AUC=1.0 for baseline images and performed better for EA morphs (AUC=0.92) than CA morphs (AUC=0.86). Therefore, humans show an ORE for morphs that compounds the difficulties associated with face-morph attack detection. Further, DCNN performance was affected by morphs; however, this did not reflect the pattern seen with human participants.