Improved privacy rotection in authentication by fingerprints

We show the feasibility of implementing a new zero leakage quantization schemeon biometric templates. In particular, we investigate the challenge that a sys-tem designer has to model the feature distributions prior to actually enrollingparticipants. For a Gaussian a priori model, we calculate the capacity and pri-vacy leakage for the commonly used biometric dataset FVC2000 and compare itto other verification schemes. We show that the zero leakage scheme achievesan equal error rate of approximately 3.5% while model mismatch leads to anaverage leakage of less than 0.1 bit per dimension. It outperforms quantizationindex modulation, both in terms of prevention of leakage and equal error rate,and binary quantization in terms of verification errors that can be tolerated.