Unidata expands liveness datasets with underrepresented faces

Unidata is adding African, Latin American and Asian participants to its liveness collection to help biometric systems train and test more evenly across demographic groups. The expansion comes as studies and NIST evaluations continue to show measurable performance gaps in presentation attack detection.

Why it matters: - Biometric systems can miss real users more often in some demographic groups than others. - Unidata is trying to close that gap by building liveness data with broader face coverage and higher-quality spoof attacks. - Better demographic balance in training and testing data can improve anti-spoofing performance in real-world use.

What happened: - Unidata expanded its liveness data collection to include additional African, Latin American and Asian participants. - The collection now covers 50 IDs and 53,870 videos. - The full attack set was recorded on the same people used in the bona fide portion of the dataset. - The dataset includes 20 IDs without silicone masks and 30 IDs with silicone masks. - Each no-mask ID includes 1,018 videos. - Each silicone-mask ID includes 1,117 videos.

The details: - The dataset combines printed 2D attacks, replays on monitors, tablets and phones, and silicone masks. - Head-movement sequences are captured on both real faces and attacks, including silicone masks. - The goal is to help developers test whether a system can detect natural movement versus attack movement. - A 2026 study of facial presentation attack detection found a 3.07 percentage-point accuracy gap between African and East Asian subjects. - NIST's 2023 evaluation of 82 passive PAD algorithms found major representation gaps in its bona fide dataset, including 15,131 images of White subjects versus 3,583 Black and 1,400 Asian subjects. - Unidata's model testing on client systems found error rates of up to 5%. - Those errors skewed toward misclassifying real users more often than presentation attacks. - The imbalance appeared unevenly across user groups.

Between the lines: - Representation alone does not solve the problem if some groups remain underrepresented in the data used to train and test a model. - The quality of the attack matters as much as the number of samples because weak spoofs can make detection look easier than it is. - Unidata's silicone masks are cast from a full plaster mold of each participant's face, which makes them harder to distinguish from the real person on camera than latex masks. - Unidata says it recruits through casting and model agencies, specialized platforms, personal contacts and international contractors to widen geography instead of repeatedly using the same pool. - Participants are told in advance that their data will be used for liveness detection training and testing, and participation is paid. - Participants can withdraw at any time. - Sessions last three to four hours and include an individual operator, a site tour and an introduction to the biometric testing process. - Annotation covers technical characteristics and metadata only, without identifying participants.

What's next: - Unidata plans to keep widening the geographic mix in its liveness collection. - The expanded dataset is positioned for developers building verification systems and anti-spoofing models. - The company is likely to use the added coverage to support more robust training and evaluation as biometric testing standards tighten.

The bottom line: - Unidata is betting that better demographic coverage and more realistic attack data will make liveness systems harder to fool and less biased in practice.