Image quality metrics can improve liveness detection
Making biometric liveness detection reliable and explainable requires recognizing that it is more than an algorithm, Neurotechnology Research and Development Engineer Mantas Kundrotas said during a presentation at last week’s Face Image Quality Workshop, hosted by The European Association for Biometrics. Quality metrics, Kundrotas argues, play a pivotal role in a complete liveness detection system.
The EAB event explored the current state of the art in face image quality metrics, and featured a series of presentations on new developments and panel discussions on ongoing issues in the field.
“Quality metrics are like a trusted virtual assistant,” Kundrotas explained towards the beginning of his presentation on “Optimizing Face Quality Metrics for Robust Liveness Detection.” They assess images, but also provide insight into their suitability. In addition to helping people understand and mitigate challenges to submitting a suitable facial image, quality metrics can help to explain the behavior of liveness algorithms, he says.
Problems with image quality can both increase the vulnerability of liveness systems and degrade their user experience, and breaking out both sides is important to improving the metrics used, according to Kundrotas, who lauds the approach taken by the U.S. National Institute of Standards and Technology (NIST) in its evaluations.
Traditional image quality metrics, however, are not comprehensive enough to explain the behavior of liveness detection systems. If users can be told why their image will not be (or was not) successfully authenticated, the user experience will be improved.
Neurotechnology has worked to optimize image quality metrics by capturing diverse quality variations by utilizing synthetic data. The company has also extracted features that explain liveness system behavior, and created adaptable machine learning models to understand the relationship between quality and liveness.
Liveness detection seems to be more sensitive to image quality than biometric matching is, Neurotechnology has found. Liveness detection is very effective for high-quality images, ineffective for low-quality images, and tends to be middling for other images, meaning that image quality scores on their own can often function as a useful proxy for confidence in liveness detection, Kundrotas says.
The feature engineering efforts can therefore inform useful feedback for real-world users to submit better-quality images for liveness detection.
Features like glasses may trip liveness detection systems with reflection, further highlighting the need for quality metrics to recognize challenging image quality considerations.
Kundrotas proceeded to review the ways models can be trained to better predict liveness behavior, and to recognize real-world image quality factors.
The company’s testing shows that metrics can be aligned with expected outcomes well enough to guide users in capturing images that will pass liveness detection, so that most users receiving feedback from a quality assistant who are unsuccessful on the first attempt will be successful by the second attempt. Those without a quality assistant tended to pass a liveness check by the third or fourth attempt.
Article Topics
biometric data quality | biometric liveness detection | biometrics | biometrics research | EAB | face biometrics | Neurotechnology
Comments