Compensating User-Specific Information with User-Independent Information in Biometric Authentication Tasks

Biometric authentication is a process of verifying an identity claim using a person's behavioral and physiological characteristics. This is in general a binary classification task because a system either accepts or rejects an identity claim. However, a biometric authentication system contains many users. By recognizing this fact, better decision can be made if user-specific information can be exploited. In this study, we propose to combine user-specific information with user-independent information such that the performance due to exploiting both information sources does not perform worse than either one and in some situations can improve significantly over either one. We show that this technique, motivated by a standard Bayesian framework, is applicable in two levels, i.e., fusion level where multiple (multimodal or intramodal) systems are involved, or, score normalization level, where only a single system is involved. The second approach can be considered a novel score normalization technique that combines both information sources. The fusion technique was tested on 32 fusion experiments whereas the normalization technique was tested on 13 single-system experiments. Both techniques that are originated from the same principal share a major advantage, i.e., due to prior knowledge as supported by experimental evidences, few or almost no free parameter are actually needed in order to employ the mentioned techniques. Previous works in this direction require at least 6 to 10 user-specific client accesses. However, in this work, as few as two user-specific client accesses are needed, hence overcoming the learning problem with extremely few user-specific client samples. Finally, but not the least, a non-exhaustive survey on the state-of-the-arts of incorporating user-specific information in biometric authentication is also presented.

Related material


EPFL authors