Cross-Device Iris Presentation Attack Detection Based on Image-Text Multimodal Alignment

Iris recognition technology, renowned for its stability and uniqueness, is extensively employed in high-security applications such as authentication and access control. However, the advent of sophisticated iris presentation attacks has emerged as a prevalent and severe security challenge. Despite numerous strategies proposed to mitigate this issue, enhancing the model's generalization across diverse devices remains a critical area for improvement. To tackle this challenge, we introduce a pioneering solution by leveraging a pre-trained CLIP (Contrastive Language-Image Pretraining) model to bolster the generalization performance in detecting iris presentation attacks. Drawing inspiration from the architecture of FLIP (Fine-tuned Face Liveness Classification), our approach further refines the loss function and presents an innovative model, denominated FLIP-Image-Text-Multimodal-Alignment. This novel framework leverages the generalization ability of image-text multimodal alignment, thereby significantly enhancing the model's capacity to generalize across various scenarios. To validate the efficacy of our proposed model, we conducted extensive experiments utilizing seven widely recognized public databases. We propose a testing protocol for assessing cross-device generalizability. The experimental results unequivocally demonstrate the superior performance of FLIP-Image-Text-Multimodal-Alignment in accurately identifying iris presentation attacks, underscoring its potential as a robust defense mechanism in the realm of biometric security.