Predicting Biometric Authentication System Performance Across Different Application Conditions: A Bootstrap Enhanced Parametric Approach

The performance of a biometric authentication system is dependent on the choice of users and the application scenario represented by the evaluation database. As a result, the system performance under different application scenarios, e.g., from cooperative user to non-cooperative scenario, from well controlled to uncontrolled one, etc, can be very different. The current solution is to build a database containing as many application scenarios as possible for the purpose of the evaluation. We propose an alternative evaluation methodology that can reuse existing databases, hence can potentially reduce the amount of data needed. This methodology relies on a novel technique that projects the distribution of scores from one operating condition to another. We argue that this can be accomplished efficiently only by modeling the genuine user and impostor score distributions for each user parametrically. The parameters of these model-specific class conditional (MSCC) distributions are found by maximum likelihood estimation. The projection from one operating condition to another is modelled by a regression function between the two conditions in the MSCC parameter space. The regression functions are trained from a small set of users and are then applied to a large database. The implication is that one only needs a small set of users with data reflecting both the reference and mismatched conditions. In both conditions, it is required that the two data sets be drawn from a population with similar demographic characteristics. The regression model is used to predict the performance for a large set of users under the mismatched condition.