3DiVi Ranks Second Among Teams in the Image Category at LivDet-Face 2026
3DiVi has achieved a top-ranking result in LivDet-Face 2026, an international competition focused on face Presentation Attack Detection (PAD). In the Image category, the company's algorithm, submitted under the identifier Anonymous0.0.2, ranked 4th out of 19 algorithm submissions. When teams are ranked by their best-performing submission, 3DiVi places 2nd overall.
In the independent evaluation, 3DiVi's algorithm achieved the following performance metrics:
ACER: 5.91% - Average Classification Error Rate, calculated as the mean of APCERavg and BPCER.
APCERavg: 3.42% - Average Attack Presentation Classification Error Rate across the evaluated attack categories, measuring the proportion of presentation attacks incorrectly classified as bona fide.
BPCER: 8.4% - Bona Fide Presentation Classification Error Rate, measuring the proportion of genuine face presentations incorrectly rejected as attacks.
Notably, 3DiVi's solution achieved 0% APCER in 6 out of 10 presentation attack categories, including display attacks, paper mask attacks, and several types of realistic skin-tone masks.
About LivDet-Face 2026
LivDet-Face 2026 is the third edition of the international Face Liveness Detection Competition, organized by Clarkson University and the University of North Carolina at Charlotte in conjunction with the International Joint Conference on Biometrics (IJCB).
The Image category featured 19 algorithm submissions from 11 participating teams. The evaluation dataset comprised 1,866 images, including 95 bona fide presentations and 1,771 presentation attack samples. An additional 375 video samples were used to evaluate submissions in the Video category.
The test data was captured using a DSLR camera and a Samsung Galaxy S24 Ultra smartphone. The evaluation covered ten categories of presentation attacks, including printed photographs, screen-based replay attacks, 3D face masks, and a newly introduced category of realistic 3D masks enhanced with skin-tone coloring and cosmetic makeup to replicate human facial appearance and texture.
All evaluations were conducted independently by the competition organizers using a sequestered test dataset that was not accessible to participants during the competition. This protocol enables an objective assessment of algorithm generalization and robustness against previously unseen presentation attacks.
3DiVi's Face Liveness Detection Technology
3DiVi's submission employs a passive, single-image face Presentation Attack Detection (PAD) approach, requiring no additional user interaction or challenge-response actions.
The algorithm is built around two complementary convolutional neural networks (CNNs), each analyzing a different spatial region of the input image.
The first network processes a tightly cropped facial region to identify fine-grained texture characteristics and localized spoofing artifacts. The second analyzes an expanded facial region to capture contextual indicators of presentation attacks, such as display or print medium boundaries, reflections, and background inconsistencies.
The outputs of both networks are fused into a single face liveness score, enabling the system to distinguish bona fide face presentations from presentation attacks.
Participation in LivDet-Face 2026 provided an opportunity to validate 3DiVi's PAD algorithm through independent testing on a sequestered dataset, covering both conventional spoofing techniques and emerging presentation attack methods.
3DiVi extends its appreciation to the LivDet-Face 2026 organizers for conducting the independent evaluation and congratulates the competition winners and all participating teams.