Spoof Detection on Ear Biometrics

Ear recognition systems are one of the biometrics-based popular person identification systems. The attacks to these biometrics identification systems become inevitable. In this context, the emerging ear recognition systems need to counter against spoof attacks. Consequently, ear anti-spoofing problem is focused in this thesis. In the biometric community, the types of attacks can be listed as; printed photo attack, display attack, replay attack, mask attack etc. In this thesis, printed photo attack is considered. Firstly, the Image Quality Assessment (IQA) methods are employed to find a solution to this problem. Therefore, 21 Full-Reference (FR) and 4 No-Reference (NR) IQA measures are implemented to extract features from ear images. In addition to this, Convolutional Neural Network (CNN) which is a deep learning method is implemented to detect impostor ear samples. Further, texture-based Binarized Statistical Image Features (BSIF) method is applied to represent the features of ear images. In this context, four different methods are proposed for distinguishing genuine ear images from impostor ones. The first one employs Decision-Level-Fusion (DLF) technique to combine FR and NR IQA measures. Secondly, three-level fusion of FR and NR IQA measures is implemented by using Score-Level-Fusion (SLF) and DLF techniques. Additionally, a CNN-based system and 5 IQA measures are combined by applying DLF technique as a third method. Finally, BSIF and CNN-based methods are fused by using DLF technique. The used databases for experiments are AMI, UBEAR, IITD, USTB Set 1, USTB Set 2 and USTB Set 3 which are publicly available. However, the spoof database for aforementioned databases is not available. Therefore, the spoof database which contains print attack is constructed in this thesis. Keywords: Ear biometrics, Spoof detection, Image quality assessment, Deep learning, CNN-based method, Texture-based method, BSIF, Printed photo attack