Presentation Attack Detection Using a Tiny Fully Convolutional Network

Abstract

Fingerprint authentication is widely used thanks to its simple process and low cost, but it is vulnerable to fake fingerprints. Many researchers have been working on presentation attack detection to ensure the security of fingerprint systems. However, the existing studies only focus on improving the detection accuracy, and let processing time and memory requirement be out of focus. Hence, it is difficult to integrate the existing algorithms to embedded and mobile systems. This paper proposes a method to detect presentation attacks using a small fully convolutional network. The proposed network is designed using the structure of the fire module of SqueezNet. The use of the fire module results in a network which has around 0.5 million parameters. The proposed network is trained using images of $32\times 32$ , $48\times 48$ , or $64\times 64$ pixels. Since the network has no fully connected layer, it can interfere with images of any size. This advantage helps to improve the detection rate and allows the proposed algorithm to be easily integrated into fingerprint systems. The experiments show an average detection error of 1.43%, which is comparable with the state-of-the-art accuracy, while the processing time and memory requirement are much reduced.