Temporal integration for continuous multimodal biometrics

Abstract

Typically, biometric systems authenticate the user at a particular moment in time, granting or denying access to resources for the complete session. This model of authentication does not appropriately address environments where a different individual may take over a system from the original user (either willingly or otherwise). We propose a multimodal system that performs authentication continuously by integrating information temporally as well as across modalities. Such continuous authentication provides ongoing (rather than onetime) verification and can easily be coupled with another system for dynamically adjusting access to privileges accordingly. We present an initial approach for temporal integration based on uncertainty propagation over time for estimating channel output distribution from recent history, and classification with uncertainty. Our method operates continuously by computing expected values as a function of time differences. Our preliminary experiments show that temporal information improves authentication accuracy. These empirical results are promising and justify further investigation. 1.