Incremental Learning-enabled Network for Fingerprint presentation attack detection

With the rapid growth of fingerprint-based identification systems, concerns about artificially generated fingerprint substantiations are rising high. To counteract these attacks, a variety of presentation attack detection (PAD) methods using robust fingerprint characteristics are being developed. However, the performance is dropped in cases when data is collected from completely new environment. In such cases, model is trained with the newly arrived fakes by re-training the existing model. Nonetheless, it has been an insignificant solution to re-train the whole system to handle the newly created spoof samples. In this research work, we propose an approach namely IncNet that is based on incremental learning and allows current PAD system to gradually learn new types of fake traits without re-training the entire model. The suggested model discriminates newly learnt fakes without sacrificing the knowledge of previously learned fakes by training deep convolutional neural network in incremental fashion. Additionally, to address the issue of data scarcity, and class imbalance problem in classification, we increased the size of data by incorporating data augmentation in our proposed model. The experiments are performed on benchmark LivDet 2013 dataset where the results demonstrate that our model exhibits promising performance with an average classification accuracy of 93.09%. Whereas, in case of augmented dataset, the model results in an increase of ~2% in average classification accuracy.