Feature Generation-Based Fingerprint Liveness Detection: A Novel Multimodal Approach

Fingerprint liveness detection is an emerging security solution that distinguishes between spoofed fingerprints and genuine fingerprints. Its importance increases because of the availability of cheap spoofing materials for fingerprints. The literature reports on fingerprint liveness detection techniques that have adopted deep learning architectures. Although multimodal traits can be an effective way to counter such spoofing mechanisms, they add hardware complexity. Hence, such multimodal traits are dropped from the usage. In contrast, this paper introduces a feature-generation procedure to exploit the benefits of multimodalities in liveness detection. Hence, a Convolutional Neural Network (CNN) architecture is proposed to generate features of iris traits when the input fingerprint image is given. This results in detecting the liveness of a fingerprint using iris features, which are generated from the input fingerprint. This averts the usage of memory to store both fingerprint and iris features on the detection side. To further effectuate the detection process, an Adaptive Focal Loss (AFL) is proposed in this paper. The proposed AFL separates the features as high-frequency and low-frequency components and performs weighted error calculation among the components. Lion algorithm trains the proposed CNN architecture to learn the features of iris traits. A chimeric dataset is constructed using the LivDet2015 fingerprint and iris datasets, in which the proposed liveness detection method is compared against the state-of-the-art unimodal liveness detection methods and multimodal liveness detection methods. The results show that the proposed liveness detection method is better than the compared methods. The results further clarify that the proposed method is less sensitive to spoofing materials and sensor types, as well as has good feature generation capability and generalization ability. An ablation study discloses the performance of AFL and the lion algorithm over traditional loss functions and optimizers in learning the fingerprint and generating the iris features.