SFincBuster: Spoofed fingerprint buster via incremental learning using leverage bagging classifier

Abstract

Listen

Translate article

The fingerprint presentation attack detection (PAD) studies are extensively explored by investigators to augment the security aspects of human authentication in biometric systems. Although existing approaches yield promising results while evaluated on data having same distribution as the training data, but their performance is declined in scenarios when data is gathered from unknown environment. In these situations model is learned with the new type of fake samples by re-training the entire model. Nevertheless, it has been a non-trivial concern to re-train the entire system to tackle the newly created fake samples as unseen artifacts are generated continuously in real-time scenarios. In this work, we expound a novel incremental learning-based approach namely; SFincBuster that can work effectively in the challenging real-time scenarios and can handle both new data as well as new observations from old classes. We train a leveraging bagging ensemble (LBE) in incremental fashion regardless of the extracted deep-level features (using pre-trained VGG19 network) being too large to accommodate into system memory, it is still possible to train our model effectively. Furthermore, the LBE integrates the simplicity of classical bagging with augmented randomization to the input and outcome of the base classifiers. To tackle the issue of change in distribution that arises with gradual changes from learned fakes to entirely new fingerprint artifacts, the SFincBuster employs LBE with ADWIN (Adaptive WINDowing) technique that continuously evaluate the performance of underlying base model and whenever a change is detected the weakest classifier is substituted with a new one. Our approach achieves high classification accuracy, even though it is not prerequisite to access all features at once. The SFincBuster is trained and evaluated on LivDet 2009, LivDet 2011, LivDet 2013, LivDet 2015 and LivDet 2021 benchmark datasets and yields maximum average classification accuracy (ACA) of 98.65% on LivDet 2013 and LivDet 2015 datasets. The model exhibits stupendous generalization capabilities with an average classification error rate (ACER) of 1.39% for known fakes (KF) and 2.84% for unknown fakes (UF). Finally, the comparable investigation perceives that the SFincBuster model demonstrates a noteworthy performance gain over the similar state-of-the-art (SOTA) approaches and achieve an improved benchmark for real-time cross-sensor, cross-material and cross-dataset scenarios.