Deep Learning-Based Fingerprint Liveness Detection Using Hybrid EfficientNetB4-ViT Architecture with Texture Enhancement

Abstract

Fingerprint-based biometric systems are widely used in security-critical applications. However, they remain vulnerable to spoofing attacks involving artificial finger replicas made from materials such as silicone, gelatin, and latex. Existing liveness detection methods based on handcrafted features often fail to generalize across varied sensors and spoof materials, necessitating more robust solutions. This paper proposes a hybrid deep learning framework that integrates EfficientNetB4 and Vision Transformer (ViT) architectures, enhanced by Local Binary Pattern (LBP) and Gabor filtering for texture preprocessing, multi-scale Convolutional Block Attention Modules (CBAM) for feature recalibration, and advanced training strategies including Focal Loss, MixUp, and CutMix augmentations. The model was evaluated on datasets from the Fingerprint Liveness Detection Competitions of 2009, 2011, 2013, and 2015, comprising over 79,000 images across diverse sensors. On a test set of 40,340 samples, it achieved an accuracy of 97.64% and an AUC of 0.9938, demonstrating that combining convolutional and transformerbased architectures effectively captures both local texture and global structural features for reliable spoof detection.