xDFPAD: Explainable Tabular Deep Learning for Fingerprint Presentation Attack Detection

Security of fingerprint authentication system is compromised when they are exposed to the Presentation Attacks (PAs). To detect these PAs under any attack conditions, an efficient Fingerprint Presentation Attack Detection (FPAD) technique is required. This work presents an explainable deep learning based FPAD (xDFPAD) method to prevent any possible PAs using any novel Presentation Attack Instrument (PAI). Our proposed method will use a tabular deep learning model called TabNet for the detection of PAs using the fingerprint quality feature based tabular data. This method combines the benefits of quality feature based FPAD solutions with the deep learning to make it white-box instead of its intrinsic block-box nature. Our proposed deep learning based TabNet model provides instance wise explainable reasoning of the output results. Performance of the method is tested on the LivDet 2017 competition database. Results indicates that our method is able to perform significantly better than the existing techniques for realistic unknown sensor or material attacks conditions. Also, our method provide the benefit of instance wise explanation of the predicted class which is not provided by the existing block-box deep learning solutions.