Impact of Face Image Quality on Morphing Attack Detection
An analysis of the impact of face image quality on MAD performance.
Benchmarks
The assessment has been carried out on the following benchmark:
- DMAD-IMARS-MQ_FULL-1.0: A dataset containing high-resolution digital face images with frontal pose, neutral expression and good illumination collected in the iMARS project [1]. The gate images have been acquired in real scenarios or in laboratory by simulating realistic conditions while the morphed images have been generated using 12 different state-of-the-art morphing algorithms and multiple morphing factors.
Testing Protocol and Performance Indicators
For each algorithm, bona fide (evaluating bona fide face images) and morph (evaluating double-identity face images) attempts are performed to compute Bona fide Sample Classification Error Rate (BSCER) and Morphing Attack Classification Error Rate (MACER). As defined in [2] the BSCER is the percentage of bona fide samples incorrectly classified as morphed samples while the MACER is the proportion of morphing attack samples incorrectly classified as bona fide samples.
From those measures, the D-EER (detection Equal-Error-Rate) is defined as the error rate at which BSCER and MACER are identical.
Different Face Image Quality Assessment (FIQA) algorithms have been used to predict the quality of images. The quality scores of enrolment and gate images are sorted in increasing order and organized in quartiles thus identifying 16 possible combinations of enrolment/gate image quality ranges (for more details see [3]).
To quantify the impact of quality, for each enrolment subset, the relative difference (percentage) in the D-EER between the gate images of bad quality (< 1st quartile) and those of good quality (> 3rd quartile) is computed:
Where D-EER(N) is the Detection Equal-Error-Rate measured on the images belonging to the N-th quartile.
Results
For each FIQA algorithm and quality quartile, the scores of three DMAD approaches are reported:
D-MAD |
Enrolment Quality |
Regressor |
CR-FIQA |
SER-FIQ |
FaceQNet |
OFIQ |
CR-FIQA |
SER-FIQ |
FaceQNet |
OFIQ |
Demorphing |
1 |
8.7% |
16.5% |
-35.2% |
-12.0% |
-97.4% |
-90.2% |
-41.6% |
-64.0% |
2 |
-30.8% |
-43.4% |
-40.8% |
-43.1% |
92.3% |
-30.7% |
-53.0% |
-47.1% |
3 |
-9.6% |
-51.6% |
-20.9% |
-13.2% |
-1.5% |
-31.3% |
-55.1% |
-41.1% |
4 |
-17.1% |
-21.2% |
-33.1% |
-50.7% |
33.8% |
49.0% |
121.8% |
-17.7% |
MagFace |
1 |
6.2% |
-10.3% |
-33.3% |
-42.8% |
-83.1% |
-61.9% |
-89.2% |
-82.5% |
2 |
-24.4% |
-49.5% |
-30.6% |
-47.1% |
9.8% |
-13.4% |
-25.7% |
-44.7% |
3 |
-6.7% |
-52.7% |
-10.3% |
12.3% |
-10.8% |
-41.9% |
-46.9% |
-33.8% |
4 |
-1.3% |
-15.4% |
2.0% |
-24.6% |
71.3% |
35.7% |
30.0% |
-39.6% |
ArcFace |
1 |
13.8% |
-0.9% |
-15.1% |
-38.3% |
-80.3% |
-41.5% |
-85.5% |
-24.4% |
2 |
14.8% |
-32.7% |
28.1% |
-43.1% |
-57.4% |
35.7% |
12.6% |
4.0% |
3 |
33.8% |
4.8% |
45.8% |
78.3% |
-26.6% |
-4.5% |
-26.6% |
-5.9% |
4 |
12.2% |
-31.6% |
-10.1% |
-16.8% |
159.9% |
33.3% |
39.1% |
-24.8% |
In addition, the impact of the quality of enrolment and gate image quality has been assessed separately:
|
Measure |
Demorphing |
MagFace |
ArcFace |
Enrolment |
Regressor |
-0.83% |
-3.07% |
24.18% |
CR-FIQA |
-2.18% |
42.28% |
43.46% |
SER-FIQ |
2.93% |
-29.82% |
-26.23% |
FaceQNet |
-10.50% |
-14.95% |
10.44% |
OFIQ |
6.09% |
-9.74% |
56.67% |
Gate |
CR-FIQA |
-13.88% |
-0.30% |
20.62% |
SER-FIQ |
-33.53% |
-34.91% |
-23.40% |
FaceQNet |
-37.80% |
-19.25% |
13.29% |
OFIQ |
-28.21% |
-30.08% |
-13.98% |
Finally, the impact that various quality components can have on D-MAD has been assessed. The following table reports the on each quality component:
|
Measure |
Demorphing |
MagFace |
ArcFace |
Quality components |
Illumination |
-12.60% |
1.95% |
15.57% |
Defocus |
-27.68% |
22.86% |
28.89% |
Pitch |
-41.11% |
-14.96% |
-1.49% |
Yaw |
-25.74% |
10.97% |
18.45% |
Shadows |
11.13% |
-31.62% |
-18.72% |
Terms and Conditions
All publications and works that cite the information reported on this web page must reference source [3].
Bibliography
[1] iMARS Project Web Site. https://imars-project.eu.
[2] ISO/IEC DIS 20059, "Methodologies to evaluate the resistance of biometric systems to morphing attacks", 2024.
[3] [2] A. Franco, M. Ferrara, C. Liu, C. Busch and D. Maltoni, "On the Impact of Face Image Quality on Morphing Attack Detection", IEEE International Joint Conference on Biometrics (IJCB), Buffalo, NY, USA, pp. 1-9, 2024.
|