# Towards Iris Presentation Attack Detection with Foundation Models

**Authors:** Juan E. Tapia, Lázaro Janier González-Soler, Christoph Busch  
**arXiv:** [2501.06312](https://arxiv.org/abs/2501.06312)  
**Date:** 2025-01-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.06312v1/

## Key points

- Foundation models have gained significant traction in artificial intelligence due to their exceptional generalization capabilities from training on vast and diverse datasets.
- Near-Infrared (NIR) Iris Presentation Attack Detection (PAD) is an area where foundation models are particularly prominent.
- Traditional databases lack subject diversity and variety of attack methods, posing a challenge in iris PAD.
- Researchers Tapia, González-Soler, and Busch propose using DinoV2 and VisualOpenClip foundation models to enhance iris PAD accuracy beyond existing deep learning methodologies.
- Utilizing foundation models can improve detection accuracy in scenarios with limited data availability compared to training systems from scratch.
- This research highlights the potential of foundation models to strengthen iris PAD systems against presentation attacks and advance biometric security measures.

## Summary

In recent years, foundation models have gained significant traction in the field of artificial intelligence. This is due to their exceptional generalization capabilities, which come from being trained on vast and diverse datasets. This trend is particularly evident in the realm of Near-Infrared (NIR) Iris Presentation Attack Detection (PAD). Traditional databases often fall short in terms of subject diversity and variety of attack methods. One key challenge in this domain is the lack of correspondence between genuine iris images and those used for attacks. These typically do not originate from the same individuals. In light of these challenges, researchers Juan E. Tapia, Lázaro Janier González-Soler, and Christoph Busch propose a novel approach to iris PAD. They leverage two prominent foundation models - DinoV2 and VisualOpenClip - to improve detection accuracy. This surpasses existing state-of-the-art performance achieved through deep learning methodologies. Notably, while systems trained from scratch have demonstrated superior results when both bona fide and attack images are available for training purposes, the utilization of foundation models presents a promising avenue for enhancing detection accuracy in scenarios with limited data availability. This research sheds light on the potential of foundation models to bolster iris PAD systems' robustness against presentation attacks. It offers valuable insights for advancing biometric security measures in real-world applications. By harnessing the power of these advanced AI models, researchers aim to address critical limitations inherent in conventional iris recognition technologies and pave the way for more resilient authentication mechanisms in diverse operational settings.