Contents I. Introduction
Over the past decades, the verification of users based on physiological biometric modalities has been the major reason for the popularity of biometrics for numerous applications [1]. Among the physiological biometric modalities, person verification using face is considered more convenient because of its ease of use and the non-intrusive nature of image acquisition. Despite impressive verification performance, and even outperforming human performance on most challenging datasets, face recognition systems still pose serious challenges when it comes to presentation attacks (PA) (i.e., spoofing attacks) [2]. A presentation attack is a deliberate attempt at impostor artifacts to impersonate the identity of genuine users by using Presentation Attack Instruments (PAIs) (according to the definitions of ISO/IEC 30107 standards [3]). With the widespread availability of facial images in the public domain, various PAIs are being created by attackers to obtain unauthorized access by presenting fake artifacts. The PAIs could be simple printed photographs or electronics display artifacts that constitute 2D presentation artifacts, while more sophisticated PAIs are 3D face mask artifacts presented in front of the Face Recognition System (FRS) to avail the access. Figure 1 illustrates the PAIs showing 2D and 3D presentation artifacts. The influence of PAIs such as 2D print, electronic display, and sophisticated 3D face masks has been studied in a substantial manner using state-of-the-art methods to demonstrate the vulnerability of facial biometrics against artifacts [4] [1], [5], [6]. Therefore, to mitigate vulnerability issues, several Presentation Attack Detection (PAD) algorithms based on handcrafted features and deep learning-based approaches have been proposed in the literature [1]. Although we note that the surveillance system operates in the visible spectrum, the majority of the face PADs employed are based on the visible spectrum [5]. On the other hand, artifacts are non-skin materials that leverage differential illumination properties compared to genuine skin across the electromagnetic spectrum because of which previous work has also shown preferences in working beyond the visible spectrum to alleviate the vulnerability of facial biometric systems [7]. More specifically, multispectral imaging has shown greater potential in this direction, thereby leveraging differential information in spatial and spectral domains. Considering these merits, in our work, we employed a multispectral imaging approach in nine narrow spectrum bands across the Visible (VIS) and Near-Infra-Red (NIR) wavelength ranges to detect presentation artifacts. Furthermore, generalizability towards unseen or unknown artifacts is a challenging task; hence, in this work, we present PAD by exploring the properties of multispectral imaging based on our newly introduced Face Presentation Attack Multispectral Database (FPAMS Database) for unseen or unknown artifacts in order to present the significance of our work. The major contributions of this work are summarized as follows: (1) Present face presentation attack detection explores the inherent properties of multispectral imaging in nine narrow bands across the VIS and NIR (530nm 1000nm) wavelength range. (2) Quantitative comparison of the image fusion (or early fusion) and score fusion (or late fusion) frameworks for face PAD. (3) Extensive experimental evaluation results are obtained on the newly introduced FPAMS database of 61650 samples, especially with the execution protocol of unseen attack detection, to confirm the performance of the proposed PAD framework.
