.png){kind=icon}
.png){kind=icon}
Return
Dual-Stream Autoencoder With Spatial and High-Frequency Feature Interaction for Contactless Fingerprint Presentation Attack Detection
.png){kind=icon}
PRE
AI
F
Y
Y
DOI：10.1109/tifs.2026.3714103.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Fingerprint presentation attack detection (FPAD) is a critical component for ensuring the security of automatic fingerprint recognition systems (AFRS). However, most contactless FPAD methods are based on supervised learning, which exhibit limited generalization capability to unknown presentation attack (PA) types and thus struggle to effectively handle unseen PAs. In addition, existing methods generally neglect the importance of high-frequency information in the FPAD task, leading to suboptimal PAD performance. To solve the above mentioned problems, this paper proposes a dual-stream autoencoder with spatial and high-frequency feature interaction, comprising a spatial reconstruction branch and a high-frequency reconstruction branch. Specifically, the high-frequency image is first extracted from the original image. Then, the original image and the high-frequency image are separately input into the spatial reconstruction branch and the high-frequency reconstruction branch, with both branches aiming to reconstruct the original image. Moreover, to learn the intrinsic correlation between the spatial and high-frequency features of bona fide fingerprints, a feature interaction module is designed to bind the spatial and high-frequency features. Furthermore, to further enhance the model’s perception of high-frequency spoof regions, a wavelet high-frequency enhancement module is introduced into the high-frequency reconstruction branch, which strengthens high-frequency detail features through multi-scale directional convolutions. Extensive experimental results demonstrate the superior performance of the proposed method on the contactless FPAD task. On the baseline CLARKSON dataset, the proposed method achieves an EER of 4.4%, which represents a 60.2% reduction compared with the current state-of-the-art method. Meanwhile, the AUC reaches 99.3%, yielding an absolute improvement of 8.5%, which fully verifies the effectiveness and superiority of the proposed method.
Keywords:
Contactless fingerprint
presentation attack detection
feature interaction
dual-stream autoencoder
high-frequency
Journal
IF:
8
Papers:
5.2K
Citations:
2.3W
Organization
Cited Papers