Fingerprint Presentation Attack Detection Using Deep Transfer Learning and DenseNet201 Network

Fingerprint presentation attack, which involves presenting spoof fingerprints to fingerprint biometric sensors to gain illicit access, is a significant challenge faced by Automatic Fingerprint Identification Systems (AFIS). As a result, various hardware-based and software-based approaches have been posited to help remedy this concern. However, the software-based methods have seen enormous utilisation relative to the hardware-based techniques due to their robust cognitive feature extraction for spoof detection. Nonetheless, most software-based techniques that utilise handcrafted features proffer shallow features for discriminating against spoofs due to their manual feature extraction process, which, as a result, affects the model s robustness. Motivated by this concern, we propose a deep transfer learning approach to automatically learn deep hierarchical semantic fingerprint features as a means of discriminating against spoofs. Experiments were conducted on the LivDet competition standard database, encompassing datasets from LivDet-2009, 2011, 2013, and 2015, resulting in the acquisition of real fingerprints and fake fingerprints fabricated from twelve (12) different spoofing materials. The developed model recorded an average classification accuracy of 99.8%, a sensitivity of 99.73% and a specificity of 99.77%, showcasing a state-of-the-art performance.