Parallel-wise global and local attention vision transformer-based generative adversarial network using fourier transform loss for generating fake iris image

Abstract

The technological progress in generative adversarial network (GAN) has facilitated the creation of highly realistic counterfeit images. This, in turn, has paved the way for presentation attacks, named as spoof attacks, targeting a range of biometric recognition systems, including those for faces, fingerprints, veins, and irises. To mitigate these risks, significant efforts have been directed toward developing presentation attack detection (PAD) methods specifically tailored to identify fake iris images generated by the technology of GAN in the field of iris recognition. However, most of them used in the existing research do not consider the difference in the frequency domain although the difference in frequency domain makes the fake iris image easier to distinguish from the true (real) one. To tackle these problems and create a more robust PAD model, we propose a parallel-wise global and local attention vision transformer-based generative adversarial network (PGLAV-GAN) that considers both global and local context, and can generate fake iris images with similar frequency domain features to true (real) ones through Fourier transform (FT) loss. The generator in PGLAV-GAN is designed to minimize the loss of feature information while allowing feature information at different resolutions to be effectively utilized, through the incorporation of nested U-shaped network (UNet++) and deep supervision as the backbone network architecture. Moreover, the performance of PAD was assessed by applying various additional post-processing techniques, confirming that the proposed method surpassed the state-of-the-art (SOTA) presentation attack (PA) methods. When the proposed model was validated through an experiment using two open databases, the average classification error rate of the PAD model was 1.51 % for the liveness detection-iris competition 2017 (LiveDetIris-2017)-Notre Dame dataset and 1.9588 % for the LiveDet-Iris-2017-Warsaw dataset. This indicates that the proposed model outperformed SOTA models, and produced fake iris images that are more authentic and sophisticated for PA, enabling more robust PAD model to be constructed with the fake iris images generated by our method and enhancing the consequent security level of iris recognition system.