Face morphing spoof attacks against biometric systems are particularly challenging for border systems to detect, and as morph attack detection systems are developed, questions have arisen about how to test such systems, and how they are affected by the quality of the images used. Progress has been made towards answers to both questions.
The iMARS (image manipulation attack resolving solutions) project held an event in Brussels and online to explore the theme “Combatting ID Fraud: New Tools for Image Manipulation Detection.” An iMARS workshop earlier this year shared some progress, but also the sobering news that morph attack incidents are increasing.
Two of the afternoon presentations for the event on Thursday were delivered by Christophe Busch and Matteo Ferrara of the University of Bologna.
Researchers with the University of Bologna have developed an evaluation platform for MAD systems, the Bologna Online Evaluation Platform (BOEP), presented earlier this year in an EAB Lunch Talk. Ferrara presented the benchmarks evaluated by the BOEP for iMARS. Thes systems are divided into single-image morph attack detection (SMAD) and differential morph attack detection (DMAD).
So far, six iMARS partners have submitted a total of 34 SMAD algorithms. The target BPCER of below 8 percent set by the iMARS project has been achieved in all SMAD benchmarks, Ferrara says, with a morph attack classification error rate (MACER) at or below 10 percent.
For DMAD, 4 iMARS partners submitted 10 algorithms. The key performance indicator in this case is BPCER of 6 percent or lower at MACER under 10 percent, and was likewise met in all benchmarks.
Ferrara also reviewed the differences between NIST’s FATE MORPH and BOEP.
Quality assessment relies on quality images
Busch discussed the importance of being able to assess biometric sample quality to the iMars project.
Large scale databases and diverse applications scenarios increase the importance of considering image quality. This helps to avoid false positives, but is also essential for interoperability. This is particular important given the many contributors to the program. They have “many different capture sites with different technology, with different levels of training,” Busch points out.
A standardized concept of what is good and what is not is the motivating concept behind the development of the face image quality assessment (FIQA). These algorithms help to predict the recognition performance that can be expected, as well as assisting with the detection of morphed images.
The EES system is set up with the ISO/IEC 19794-5 in mind, so that provides some requirements, and the 29794-5 standard, which OFIQ (Open Source Face Image Quality) is a reference implementation for, provides measurements for them. They must be applied to pre-enrolled reference images, live-enrolled reference images, like at an EES kiosk, and probe images, such as from a biometric automated border control gate.
The unified quality score (UQS) specified by ISO/IEC 29794-1 provides the overall assessment, and defect measurement provides explainability. OFIQ software provides 28 quality measures, starting with the unified quality score, covering capture device and subject variables.
Numerous algorithms have been proposed, and the iMARS project determined three criteria for assessing the proposals: accuracy, low computational complexity and the liberality of the license.
Images with OFIQ-UQS above a certain threshold, perhaps around 70, will be useful enough to store and use.
The success of the metric is shown in the performance improvements seen in matching performance as low-quality images are discarded, Busch explains, providing an answer to one of the central questions of the iMARS project; “How can we measure the impact of face image quality on biometric recognition performance.”
Busch reviewed the quality components that go into the unified score, and how they are measured based on the position of facial landmarks and other observable characteristics. Expression neutrality has proved relatively complex, but not impossible to measure.
Research published in September shows that a positive correlation between images assessed as high-quality and success in morphing attack detection (MAD).
Future work will focus on demographic variability and adding missing quality components, such as motion blurring.
Article Topics
biometrics research | Bologna Online Evaluation Platform (BOEP) | face morphing | iMARS | Morphing Attack Detection (MAD) | Open Source Face Image Quality (OFIQ) | spoof detection
