Page - 69 - in Joint Austrian Computer Vision and Robotics Workshop 2020

ArtefactType aGen aImp Baseline 0.2346 0.1257 Wax 0.1222 0.1236 Wax traced 0.1199 0.1220 WaxCLAHE 0.1252 0.1199 Silicone 0.1285 0.1297 Silicone traced 0.1250 0.1250 SiliconeCLAHE 0.1274 0.1283 Table 2. Average genuine (aGen) and impostor (aImp) FV comparison scores obtained when verifying bona fide samples only (baseline) compared to verifying bona fide samplesagainstPAsusingdifferentartefact typesfor light transmissionFVrecognition. images have been used. The scores have been av- eraged over all three fingers and paper types for il- lustration purposes because of the small variation in their scores. It is immediately noticeable that none of the spoofing artefacts is meeting the quality re- quirements, since theobtainedgenuineand impostor scores are not differentiable. This is also true for the visually promising traced wax artefacts. Therefore, a further refinement of these artefacts is necessary to come up with a dataset of sufficient quality as re- quired for a sensiblePADevaluation. 5.3.Results: PADPerformance Following the evaluation of the produced spoof- ingartefacts’quality, thissectioncoversthedetection performance of the PAD system described in section 4. The evaluation of the PAD system is only per- formed for presentation attacks using HV artefacts due to insufficient quality of the FV artefacts. The available genuine and spoofed data was split 50/50 onauserbasis for trainingand testing. ThePADdetectionperformanceunderdifferent il- lumination conditions in terms of D-EER, BPCER @ APCER<=0.001 (BPCER1000) and BPCER @ APCER=0 (BPCER0) is reported in Table 3. It be- comes apparent that the artefacts are harder to de- tect under 950nm NIR than under 850nm one. This might be due to varying reflectivity and absorption propertiesof theveinpatternprints fordifferentNIR wavelengths. The PAD system has some problems in correctly classifying the palmar 950nm artefacts, however the PAD performance can be considered good to excellent acrossallHVartefacts. D-EER BPCER1000 BPCER0 Dorsal850 0.22 0.43 0.43 Dorsal950 0.33 0.65 0.65 Palmar850 0.00 0.00 0.00 Palmar950 6.04 30.43 30.43 Table 3. Performance values (in %) for hand veins PAD evaluation 6.Conclusion Presentation attacks are still a major problem in many applications of biometric recognition systems. Recent publications have shown that even vascular pattern based systems are susceptible to this kind of attack. In this work, we investigated two ap- proaches to produce presentation attack artefacts, one for finger veins and one for hand veins. We also developed a suitable presentation attack detec- tion scheme for vein recognition systems based on a natural scene statistics framework. We established a hand vein presentation attack dataset, consisting of 100 presentation attack samples and the correspond- ing original samples, which is publicly available as partof thePROTECTMMDBv21. The PAD evaluation results on the established dataset showed that the proposed PAD approach achieves a good performance in detecting the fake representations. The verification experiment further revealed that if the fake representations are not de- tected, they achieve a rather high verification rate, i.e. that there is a good chance that a presentation attack is successful if no suitable PAD approach is employed. Ourfutureworkwill includetestswithother types ofpresentationattackartefacts for thehandveinsand the establishing of a presentation attack dataset for fingerveinsaswell. References [1] A.P.S.Bhogal,D.Söllinger,P.Trung,J.Hämmerle- Uhl, and A. Uhl. Non-reference image quality as- sessment for fingervein presentation attack detec- tion. In Scandinavian Conference on Image Ana- lysis. Springer, 2017. [2] H. Hofbauer and A. Uhl. Applicability of no- reference visual quality indices for visual security assessment. In Proceedings of the 6th ACM Work- shop on Information Hiding and Multimedia Secur- ity, 2018. 1Willbe releasedathttp://projectprotect.eu 69