Page - 66 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Section 4 presents our proposed PAD approach. The experimental evaluation is described in Section 5. Section 6 concludes this paper and gives an outlook on futurework. 2.RelatedWork Finger- and hand-veins have been shown to be susceptible to spoofing [26, 24]. PAD approaches help indetectingpresentationattacksandcanbecat- egorised into liveness-based (rely on signs of vital- ity, e.g. capturing the heartbeat), motion-based (ana- lysemovementsduring thecapturingprocessand try to detect unnatural ones) and texture-based meth- ods (detect and analyse textural artefacts present in the image). While the first two categories require a videoorasequenceofconsecutive images tobecap- tured, texture-basedmethodscanbeapplied tosingle images. One liveness based approach is presented in [19], which applies motion magnification tech- niques. The majority of the proposed PAD schemes are texture-based ones, e.g. a Fourier, Haar and Daubechies wavelet transform based one [16], ex- ploiting differences in the bandwidth of vertical en- ergy signals. A binarised statistical image features based one and some others based on Riesz trans- form, localbinarypatterns (LBP), localphasequant- isation and Weber local descriptors are presented in [25]. Another approach [23] uses a windowed dynamic mode decomposition (W-DMD) to detect spoofed finger vein images. Even baseline LBP [20] and some LBP variants and extensions of LBP [10] provedtobeeffectivefor thetaskoffingerveinPAD. Several other approaches are utilising image quality assessment methods (IQA), e.g. [15] and [1] which detection accuracy turns out to be highly dependent on the particular dataset. In [22] the authors showed that the classification accuracy can be improved by incorporating natural scene statistics (NSS) [13]. A very different approach for PAD detection is to use a photo-response non-uniformity (PRNU) technique to differentiate PA data from genuine samples [12]. Furthermore, a CNN-based approach has been pro- posed in [17]. 3.PresentationAttackApproaches Capturing the vein pattern using an appropriate capturing device forms the basis of vein recognition ingeneralandfinger-andhand-veinPAevaluationin particular. Therefore, we utilise the PLUSVein fin- gerveinscanner [7]andthePLUShandveinscanner (a) Genuine (b) Post-processed (c) Recaptured (a) Genuine (b) Post-processed (c) Recaptured Figure1.HandveinPAartefactsfor950nmreflectedlight illumination captured with the PLUS hand vein scanner [8]: genuine image (a), post-processed image for printing (b)and re-acquired printed image (c). [8] as capturing devices to prepare our finger- and hand-vein spoofing artefacts as well as for recaptur- ing the artefacts. The interested reader is referred to the authors original publications for more details about those capturing devices. In the following, the productionof thehandandfingerveinspoofingarte- facts is described. These spoofing artefacts are then again presented to the capturing devices mentioned above. 3.1.HandVeinSpoofingArtefacts The hand vein capturing device is used to acquire reflected light images in two different wavelengths (850 and 950 nm). Since printouts of finger vein patterns have shown to yield successful presentation attacks [26], we decided for this approach as an at- tackscenario for thehandveinrecognitionsystemas well. Our spoofing attack samples are derived from samples contained in the publicly available PRO- TECTVein dataset, which is part of the PROTECT MultimodalBiometricDatabase [21]. Thehandveinspoofingattacksamplesarecreated byfirstselecting100imagesbasedonthevisibilityof the vein pattern (5 dorsal and 5 palmar for one hand of 10 users). Afterwards, a region of interest (ROI) is manually cropped from the images. These ROIs arethenpost-processedusingaContrastLimitedAd- aptive Histogram Equalisation (CLAHE) and Gauss filtering, toenhance thevisibilityof thevascularpat- tern and remove the skin texture and hair to even- tually obtain smooth images. Afterwards, the post- processed images are scaled to approximately match 66