Fig 1.
Illustration of morphed face images created using different morphing approaches.
The images on the left and on the right represent the corresponding two bona fide face images. Face images are republished from [6] under a CC BY license, with permission from Prof. Karl Ricanek Jr, University of North Carolina at Wilmington, original copyright 2006.
Fig 2.
General workflow of our proposed pipeline for image pre-selection.
Embeddings were extracted from one sample of each subject. Distances between embeddings were calculated. Faces were paired based on a low distance between embeddings. Pairs were then morphed, and morphed images were verified against bona fide probe images. Furthermore, Morphing Attack Detection has been conducted. The image pre-selection steps are further illustrated in Algorithm 1. The processing steps were performed using different FRSs and different morphing algorithms. Face images are republished from [6] under a CC BY license, with permission from Prof. Karl Ricanek Jr, University of North Carolina at Wilmington, original copyright 2006.
Table 1.
The number of embeddings per FRS.
Table 2.
The verification thresholds on the unnormalized Cosine distances for each open-source FRS.
Thresholds were calculated on the UNCW data set. The corresponding FNMRs are illustrated next to the thresholds as decimal fractions.
Fig 3.
Morphing Attack Potential (MAP).
The MAP is a matrix describing the success of a data set of morphed images to fool a set of FRSs using multiple attack attempts. Several FRSs (x-axis) are attacked with several mated Morphing Attack attempts (y-axis). The element of a MAP matrix describes the proportion of successful verifications of both attackers (i.e., both contributing subjects of each morph) at a given number of attempts (i.e., number of different bona fide images for both subjects) and with a particular number of fooled FRSs. Note that MAP was calculated as a decimal fraction within the range [0;1].
Fig 4.
ArcFace or MagFace embeddings were extracted from bona fide images and morphed images. Differential embeddings have been created by subtraction of either the embeddings of a bona fide image from a morphed image or by the subtraction of a bona fide image from a different bona fide image of the same data subject. The differential vectors have been re-scaled to N(0, 1). A classifier was trained (on ArcFace and MagFace differential embeddings, separately) to differentiate between bona fide images and morphed images. Face images are republished from [6] under a CC BY license, with permission from Prof. Karl Ricanek Jr, University of North Carolina at Wilmington, original copyright 2006.
Fig 5.
Mated morphs comparison success rates for different image pre-selection embeddings.
prodAvgMMPMRs (y-axes) are plotted for different pre-selection methods (x-axis & color-coded). Density is plotted in horizontal direction. Median values are illustrated by horizontal black bars. The same pairs were morphed by different morphing methods (rows). Random assignments of the morphing pairs are displayed in the left-most column. All morphs were verified using ArcFace and MagFace (columns). See S2 Fig for verifications using DeepFace and VGG-Face. Note that prodAvgMMPMR was calculated as a decimal fraction within the range [0;1].
Fig 6.
Morphing Attack Potential (MAP) of morphs generated by the UBO morpher.
Different FRSs were used for image pre-selection, i.e. ArcFace, DeepFace, VGG-Face, or MagFace (different heatmaps). Alternatively, pairs were randomly assigned (bottom heatmap). For each FRS used for pre-selection, the resulting morphs were verified against four bona fide images of each subject. The ratio of successful attempts for both subjects is shown on each y-axis of each plot. In addition, different FRSs were used to verify the paired morphs, four open-source FRSs and two COTS FRSs. The percentage of successful attacks across multiple FRSs is plotted on each x-axis. The MAP is shown and color-coded in each cell and describes the proportion of successful verifications for a given number of attempts (y-axes) and FRSs (x-axes). Note that the MAP was calculated as a decimal fraction in the range [0;1].
Fig 7.
ECDFs for distance scores of the open-source FRSs.
Mated, non-mated, and mated morph comparisons were performed. Morphs were created using the UBO morpher. The distance values for the comparisons are shown on the x-axis. The (cumulative) proportion of positive verifications at a certain distance score is plotted on the y-axes. Different FRSs were used for verification (rows). The different types of comparisons are color-coded, i.e., mated, non-mated, or mated morph comparisons, including morphs pre-selected with the help of face embeddings of the different FRSs. The dotted vertical lines indicate the 0.1% FMR threshold for each FRS used for verification.
Fig 8.
ECDFs for similarity scores of the COTS FRSs.
Mated, non-mated, and mated morph comparisons were performed. Morphs were generated using the UBO morpher. The different similarity scores for the comparisons are displayed on the x-axis. The (cumulative) proportion of successful verifications at a particular similarity score is plotted at the y-axes. Note that because similarities instead of distances were used, the interpretation of the x-axes must be flipped compared to Fig 7. Different COTS FRSs were used for verification (rows). The different types of comparisons are color-coded, i.e., mated, non-mated, or mated morph comparisons, with morphs pre-selected with the help of face embeddings of certain FRSs. The dotted vertical lines indicate the 0.1% FMR threshold for each FRS used for verification.
Table 3.
Relative Morph Match Rates (RMMRs).
Images were morphed using different morphing algorithms, pre-selected using embeddings of different FRSs or alternatively, randomly pre-selected, and verified using different FRSs. The RMMR corrects the MMPMR by the FNMR of the verification FRS (see Eq 4). The highest values row-wise are highlighted in bold, leaving out the quasi-diagonal elements, i.e., if pre-selection and verification FRSs coincided. Note that RMMR was calculated as a decimal fraction within the range [0;1].
Table 4.
Average ranks for RMMR values for the different pre-selection methods.
Pre-selection was either performed using random assignment of pairs or based on embeddings of four different FRSs.
Fig 9.
BPCER10 values of the classifiers tested on differently morphed and differently pre-selected data sets are shown. Left: metrics from a D-MAD algorithm trained on ArcFace embeddings. Right: Metrics from a D-MAD algorithm trained on MagFace embeddings. The images morphed by different morphing algorithms are shown in different colors. The pre-selection methods to generate the pairs for morphing are distributed along the x-axes. Note that BPCER10 was calculated as a decimal fraction within the range [0;1].
Fig 10.
DET curves of the D-MAD approaches.
Left column: D-MAD approach using ArcFace embeddings (original version). Right column: D-MAD approach using MagFace embeddings. Morphs of the different morphing algorithms are separated by rows. Data subsets of differently pre-selected morph pairs are color-coded. The BPCER is plotted against the MACER. Dotted lines indicate the positions where BPCER or MACER are 0.1 (i.e., 10%) and 0.05 (i.e., 5%). Note that both rates were calculated as decimal fractions within the range [0;1].