PAD for ID cards improving (gradually): IJCB 2025 competition

Presentation attack detection on ID cards is improving, although the technology is still facing challenges, according to the newly published results of the second PAD-ID Card 2025 competition.

The competition was held within the International Joint Conference of Biometrics (IJCB) 2025. It was organized by the Darmstadt University of Applied Sciences (Hochschule Darmstadt, h-da), Facephi and Fraunhofer Institute for Computer Graphics Research (Fraunhofer-IGD) and supported by the EU’s Horizon 2020 program and German government agencies.

The winners of the PAD on ID cards competition will be presented during IJCB 2025 from September 8 to 11 in Osaka, Japan. The competition follows up on the inaugural edition held last year.

This year’s competition evaluated 74 submitted models from 20 teams, with two winning teams declared in two tracks. Aside from the two tracks, the competition included other new elements, including an automatic evaluation platform for automatic benchmarking and a new ID card dataset for Track 1 teams, which served as the baseline dataset for the training and optimization.

The first place in Track 1 was awarded to the “Dragons” team with an average ranking of 40.48 percent and Equal Error Rate (EER) of 11.34 percent.

During the Track 1 challenge, participants were invited to train or retrain their models on a shared synthetic dataset designed to simulate real-world ID verification scenarios and identify the best algorithm based on a common dataset.

The dataset was generated using digitally generated ID card templates that mimic a wide range of identity documents from several countries.

In Track 2, the “Incode” team achieved the best results with an average ranking of 14.76 percent and 6.36 percent EER. The team was the only one that surpassed the baseline on Track 2 and improved on the results from the PAD-IDCard-2024 competition.

The main challenge of Track 2 was to identify how the number of bona fide images, countries, and diversity of attacks are relevant to obtain generalization capabilities. The participants were allowed to train their models on any open set dataset available on the state of the art and complement their data with private datasets.

“These results suggest that PAD on ID cards is improving, but it is still a challenging problem related to the number of images, especially of bona fide images,” notes the conference paper summarizing the results.

The competition opens the debate on obtaining and increasing the number of images and countries while complying with privacy rules.

“The focus of this edition was to highlight the algorithms and dataset sizes,” the paper adds. “However, the efficiency of the model must also be considered in future editions, because many foundation models can obtain very good results, but at the cost of an expensive consumption of resources.”

The same problem is also being addressed by the RIVR competition held in the U.S. by DHS S&T, with similar findings last year.

Article Topics

digital ID  |  document liveness  |  identity document  |  IJCB  |  presentation attack detection  |  research and development