OCT Fingerprint Presentation Attack Detection Based on Dual-Branch Reconstruction Differences

The emergence of optical coherence tomography (OCT) as a noninvasive imaging technique has advanced research in automated fingerprint recognition systems (AFRSs). Several methods leveraging OCT for fingerprint presentation attack detection (PAD) have been proposed, offering promising solutions to enhance the security and reliability of biometric authentication systems. Effectively detecting unknown presentation attacks (PAs), however, remains a challenging issue. A reconstruction-based strategy that requires only bonafide data for training could be a solution to address the issue of data dependency. It detects unknown PAs by computing differences between the input image and the reconstructed image. Two key challenges are, nevertheless, presented in reconstruction-based PAD: style differences introduced during reconstruction and inadequate reconstruction performance when dealing with PA samples, which affect detection performance. In order to tackle these issues, this study proposed a novel dual-branch reconstruction differences-based PAD (RePAD) method for fingerprints captured by OCT. The method consists of two distinct branches: the Recovery branch and the Equality branch. The Recovery branch is specifically designed to enhance the network’s ability to reconstruct any inputs into bonafide-like images. A simulated anomaly module is further introduced to generate abnormal images for participation in network training without any preknowledge of PAs. Meanwhile, the Equality branch and the style difference loss function work together to eliminate any potential bias introduced by variations in image style. The antispoofing score is then calculated based on the reconstruction differences of two branches in terms of B-scan wise and instance wise. This architecture can effectively expand the decision boundary for PAD by enabling more robust differentiation between bonafides and PAs. The proposed method was evaluated using the ZJUT-EIFD public dataset. The results demonstrate that the proposed approach achieves the best performance in terms of both B-scan-wise and instance-wise PAD. Comparative analysis and ablation experiments reveal that the dual-branch structure effectively enhances the reconstruction ability and mitigates style differences caused by reconstruction.