Cross-sensor Generalization for Fingerprint Presentation Attack Detection Leveraging Local Feature Enhancement

The widespread use of fingerprint recognition systems has heightened the need for robust presentation attack detection (PAD) to safeguard against spoofing attempts. Despite the success of convolutional neural network (CNN)-based PAD methods, they are constrained by significant information loss and limited generalization capabilities against unseen materials and different sensor data. In this work, we present a fingerprint PAD framework that integrates three core strategies to address the challenges posed by cross-sensor setting. First, we introduce a sensor scaling factor as a preprocessing mechanism to ensure consistent level of fingerprint information across different sensor data during training and testing. Secondly, we introduce a network architecture that utilizes a liveness score predictor to enhance relevant local features while maintaining global context. Finally, we introduce a training procedure that leverage receptive field-wise error calculation which enables the network to learn important features from every region of the fingerprint image and also mitigates the overfitting issue by promoting label density during training. Through this integrated design, our method effectively combines global context and localized discriminative cues, leading to substantial gains in both cross-material and crosssensor generalization. Experimental evaluations demonstrate that our proposed method achieves substantial improvements in crossmaterial generalization, with enhancements of 59% and 8.2% on the LivDet 2017 and LivDet 2015 datasets, respectively, compared to the current state-of-the-art methods. Moreover, the proposed method substantially outperforms existing approaches in crosssensor generalization, achieving an average classification error (ACE) of 4.26% on LivDet 2017 and 1.75% on LivDet 2015, representing improvements of 7.6% and 35%, respectively, over prior competitive methods.