Template Protection for PCA - LDA - based 3 D Face Recognition System

Nowadays systems for authentication are mainly based on passwords and possession. This implies some problems: For high security purposes long passwords are required. Because of the fact that more and more passwords are needed for several applications it gets harder to remember all of them. So what usually happens is that only a few number of weak passwords are used for all of these applications. Biometrics is a way out of this trend, as biometric features can not be easily stolen or forgotten. It offers a higher level of convenience – two main concerns using biometric system consist of security and privacy issues, it is hard to prove the security of such a system. If biometric templates get stolen some problems arise: A potential identity theft could be one consequence. Merging the data from different biometric databases implies being able to built profiles about the users, to cross-match the templates. In addition this modality is compromised forever, since there is no possibility of re-enroling the biometric. In general biometrics bear potential information about medical states, all these facts lead to a low public acceptance. Template protection techniques are invented to overcome these problems – the basic idea is to combine biometrics with cryptography. The captured samples are not stored directly but in a transformed secure form – a random number is merged with the sample. In this way, different secure templates can be created as often as needed from the same modality, which implies a solution of the mentioned problems. In this thesis face recognition is chosen to be combined with template protection, since the facts are that face images are easy to capture and that its usage as a biometric has a long tradition and a high user acceptance. Problems of the publicity and quality concerning 2D images can be solved using 3D facial scans instead. Principle component analysis (PCA) and linear discriminant analysis (LDA) are used for feature extraction as these algorithms are gold standard for 2D face recognition. The mentioned algorithms are implemented; evaluation is done using the FRGC v2.0 range database. Different techniques for spike removal and face region selection are used to enhance the source material. It is seen that recognition rates for PCA/LDA algorithms strongly depend on the quality of this material and the chosen training sets, good results could already be achieved using small secret sizes.