Large deviations performance analysis for biometrics recognition

In the selection of biometrics for use in a recognition system and in the subsequent design of the system, the predicted performance is a key consideration. Most information available to engineers about biometrics today is in the form of noisy images or signals. A typical challenge here is to select a robust set of deterministic features to guarantee a good recognition performance. An alternative to feature extraction approach is to model the raw data or a set of features as a realization of a stochastic process. While for physical signatures estimation of a stochastic model is more straightforward, for many biometric signatures it remains unapproachable challenge. In this work, we proceed with an analysis assuming the stochastic models are known, providing the analytical tools that will ultimately predict performance when these distributions are better characterized. Major challenges of biometric system design today include biometric fusion, adjustment of thresholds in decision functions with the goal to achieve an optimal performance, collection of data under conditions of environmental and demographic variations for the purpose of system testing and performance prediction [2,3,5]. In this work, we address two of them. They are performance prediction and capacity for biometrics based identification systems. For our analysis, we model templates stored in the database of an identification system as noisy realizations of independent an