A performance analysis of logistic regression and support vector machine classifiers for spoof fingerprint detection

The unique and rich details present in human fingerprints has nowadays made fingerprint verification common in a number of authentication systems. However, fraudulent fingerprints made in order to imitate the live ones with a view to deceiving these systems are now common. It is therefore important to develop a robust authentication system capable of detecting or recognizing liveness in any fingerprints presented to the authentication system. This paper therefore developed a software-based fingerprint spoof detection model using deep convolutional neural network (DCNN) features extracted from a set of live and spoof fingerprint images. The extracted features are then separately fed to two trainable classifiers namely logistic regression (LR) and support vector machine (SVM). Three variants of the SVM were used based on the training kernel: radial basis function (RBF), linear and polynomial kernels). Experiments performed on the Biometrika datasets obtained from the LivDet2009 database yielded a true accept rate (TAR) of 97.96%, a true reject rate (TRR) of 85.95%, and an average classification accuracy (ACA) of 91.94% for the LR classifier. An accuracy of 100% was obtained for the SVM trained using RBF kernel while an accuracy of 96.16% was obtained using the linear kernel. SVM trained using polynomial kernels of orders 2 and 3 yielded accuracies of 99.6163% and 99.4244% respectively. The model's accuracy was observed to decrease as the polynomials order increased. All experiments were carried out in MATLAB programming environment.