Learning sparse kernel machines with biometric similarity functions for identity recognition

We investigate the application of similarity-based classification to biometric recognition, interpreting similarity functions used in biometric systems (i.e., matching algorithms) as kernel functions. This leads us to formulate biometric recognition as a distinct two-class classification problem for each client, which can be solved even when no representation of biometric samples in a feature space of fixed dimensionality is available. We discuss the relationship of our approach with cohort-based methods, and show that using support vector machines exhibits several advantages, in terms of the automatic selection of the cohort size and elements, and of the possible update of each user model. A biometric verification setting is considered for the formulation of the approach, but experimental results with face and fingerprint data sets are reported for both verification and identification settings.