Related Experiment Video
Updated: Jul 19, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.9K
LFLDNet: Lightweight Fingerprint Liveness Detection Based on ResNet and Transformer
Kang Zhang1, Shu Huang1, Eryun Liu2
1Engineering Research Centre of Molecular & Neuro Imaging of the Ministry of Education, School of Life Science and Technology, Xidian University, Xi'an 710071, China.
Sensors (Basel, Switzerland)
|August 12, 2023
Summary
A new lightweight network enhances fingerprint liveness detection against spoofing. It uses CycleGAN for generalization and a ResNet with self-attention for improved performance and reduced complexity, achieving high accuracy on benchmark datasets.
Area of Science:
- Biometrics and Security
- Artificial Intelligence
- Computer Vision
Background:
- Fingerprint recognition systems are advancing rapidly, making liveness detection crucial for preventing spoofing attacks.
- Convolutional neural networks show promise for liveness detection but require improvements in generalization to unknown materials and computational efficiency.
- Existing methods struggle with the generalization ability for unknown materials and the computational complexity of deep learning models.
Purpose of the Study:
- To propose a novel, lightweight network for robust fingerprint liveness detection.
- To enhance the generalization capability of models against spoofing attempts using diverse materials.
- To improve detection performance while reducing computational overhead in fingerprint identification systems.
Main Methods:
- A novel lightweight network combining foreground extraction, fingerprint image blocking, CycleGAN-based style transfer, and an improved ResNet with a multi-head self-attention (MHSA) mechanism.
- CycleGAN was employed to improve the model's generalization ability for fake fingerprints derived from unknown materials.
- An improved ResNet incorporating a Transformer with MHSA was utilized to enhance detection performance and decrease computational load.
Main Results:
- The proposed method demonstrated effective Region of Interest (ROI) extraction and an end-to-end data structure, increasing data volume.
- Experiments on LivDet2011, LivDet2013, and LivDet2015 datasets showed competitive results, with an average classification error of 1.72% on LivDet2015 across all sensors.
- The network achieved 95.27% accuracy on small-area fingerprints with a significantly reduced parameter count (0.83 M).
Conclusions:
- The developed lightweight network effectively distinguishes real from fake fingerprints, addressing limitations of previous deep learning approaches.
- The integration of CycleGAN and MHSA-based ResNet significantly enhances model generalization and detection accuracy while minimizing computational resources.
- The proposed method offers a promising solution for secure and efficient fingerprint liveness detection, particularly for challenging scenarios involving unknown spoofing materials.
