GRU-AUNet: A Domain Adaptation Framework for Contactless Fingerprint Presentation Attack Detection

TL;DR

The paper addresses the challenge of robust presentation attack detection for contactless fingerprints, where traditional domain-adaptation methods struggle to generalize to unseen attacks. It introduces GRU-AUNet, a Swin Transformer–based UNet with a Dynamic Filter Network bottleneck and a GRU-enhanced attention path, trained with a combined $L_{FC} = L_{Focal} + \lambda L_{Contrastive}$ to discriminatively separate live and spoof fingerprints. The approach achieves state-of-the-art cross-dataset performance on CLARKSON, COLFISPOOF, and IIITD, reporting metrics such as $APCER = 1.2\%$ and $BPCER = 0.09\%$ on CLARKSON, and demonstrates strong generalization in cross-dataset and 5-fold cross-validation scenarios. These results indicate that the proposed architecture offers robust, scalable, and generalizable PAD for contactless fingerprint systems, with significant practical implications for security in biometric authentication.

Abstract

Although contactless fingerprints offer user comfort, they are more vulnerable to spoofing. The current solution for anti-spoofing in the area of contactless fingerprints relies on domain adaptation learning, limiting their generalization and scalability. To address these limitations, we introduce GRU-AUNet, a domain adaptation approach that integrates a Swin Transformer-based UNet architecture with GRU-enhanced attention mechanisms, a Dynamic Filter Network in the bottleneck, and a combined Focal and Contrastive Loss function. Trained in both genuine and spoof fingerprint images, GRU-AUNet demonstrates robust resilience against presentation attacks, achieving an average BPCER of 0.09\% and APCER of 1.2\% in the CLARKSON, COLFISPOOF, and IIITD datasets, outperforming state-of-the-art domain adaptation methods.

Figures (3)

• Figure 1: Different spoofed samples from CLARKSON ,COLFISPOOD, and III-TD datasets were employed in this paper. First four spoof samples are related to CLARKSON dataset (WOODGLUE,PLAYDOH, PHOTOPAPER, ECOFLEX), and others are spoof samples from COLFISPOOF dataset, indoor and outdoor photopaper are realted to III-TD dataset.
• Figure 2: GRU-AUNet Architecture: (a) The encoder processes $256 \times 256$ RGB inputs, dividing them into patches and tokenizing into $C$-dimensional vectors. (b) These vectors pass through Swin Transformer blocks, patch merging layers, and the Dynamic Filter Network in the bottleneck, which dynamically adapts filter responses. (c) The decoder mirrors this structure, upsampling features to recover spatial resolution for anti-spoofing detection. (d) A attention path replaces skip connections, focusing on critical features to improve classification accuracy. (e) In the bottleneck, the Dynamic Filter Network processes input feature maps through spatial and channel filter branches. The spatial branch applies a $1 \times 1$ convolution, while the channel branch uses global average pooling, fully connected layers, and ReLU activations.
• Figure 3: Comparison of spoofing detection model performance using different loss functions, with APCER (attack samples misclassified as genuine) on the x-axis and BPCER (genuine samples misclassified as attacks) on the y-axis.