Experimental investigation of OC-SVM for multibiometric score fusion

Different single biometric systems carry in their outputs redundant and complementary information. The concatenation of match scores from various systems in a single feature vector to feed the classifiers can provide an opportunity to develop more efficient system compared to other fusion schemes. In this work, we investigate the performance of classifier based biometric score fusion. For this purpose, the One-Class SVM (OC-SVM) classifier is employed since, in the general case of biometric systems, the data are highly unbalanced or available from only a single class. Experiments are conducted on the well known NIST-multimodal partition of the BSSRI database and results are reported using genuine acceptance and false acceptance criteria. The obtained results show the effectiveness of the OC-SVM compared to the standard two-class SVM classifier as well as to other score fusion schemes.