Mitek’s passive liveness impresses in biometric image benchmark

Results are beginning to emerge from the latest challenge in the LivDet biometric liveness detection competition series, with Mitek Systems revealing strong results in the category that most closely matches real-world applications of its technology.

The company reports that its biometric facial liveness detection algorithms produced the three highest results in the image category of the 2026 LivDet-Face competition.

The Center for Identification Technology Research (CITeR) organized the test in collaboration with Clarkson University and the University of North Carolina at Charlotte.

In the image category, organizers ran submitted biometric algorithms against a collection of bona fide and attack images that had not been disclosed to participants.

This tests if a presentation attack detection (PAD) algorithm can recognize attacks outside the data available to its developer.

Mitek describes its technology as passive liveness detection, and emphasizes the benefit of avoiding introducing friction with fraud prevention technologies.

“These results reinforce our commitment to deliver AI-powered identity verification that helps organizations strengthen security while providing seamless digital experiences,” says Garrett Gafke, chief operating officer at Mitek.

Mitek says it built its liveness detection with an ensemble of models trained using proprietary datasets that included demographic and device diversity and more than 20 presentation attack types.

Participating companies had to submit their PAD software into Docker containers in PNG or JPEG files. Organizers executed the containers on servers at Clarkson University. The environment included Intel Xeon processors, an Nvidia RTX 4090 GPU, and 128GB of memory.

Internet connections and calls to external services were prohibited to prevent participating companies from sending test images to their cloud infrastructure.

Organizers decided that scores of 50 or higher were treated as bona fide, while lower scores were classified as attacks.

Mitek says the withheld material exposed the algorithms to printed faces, screen-based replays, and sophisticated 3D masks, including attack conditions the entrants had not previously encountered.

The company did not disclose the details for the models associated with the three leading positions. Some differences could be architecture, training data, ensemble composition, and other experimental settings.

The company says its passive PAD technology combines multiple AI models trained on proprietary data.

The 2026 claim follows several earlier evaluations that originated with ID R&D.

Mitek acquired the biometric software developer in 2021 and subsequently brought its facial PAD technology under the Mitek brand.

IDLive Face had already appeared in the LivDet-Face 2024, where it was identified as an entrant in the image category and finished second.

It had an average classification error rate of 5.53 percent against 4.93 percent for the anonymously submitted first-place algorithm.

The newly claimed results suggest that Mitek’s technology remained competitive when LivDet-Face returned in 2026.

In a separate U.S. government assessment, there are more details on the performance of Mitek’s algorithm.

The company revealed that its software (labeled PAD-P9) produced an APCER of zero, which means none of the attacks used in the evaluation passed its liveness check.

The available information about LivDet-Face 2026 does not yet show how the three Mitek entries performed in absolute terms.

Article Topics

biometric liveness detection  |  biometrics  |  CITeR  |  face biometrics  |  LivDet  |  Mitek  |  presentation attack detection