On the predictability of biometric honey templates, based on Bayesian inference

In high level security environments, data protection and leakage prevention remains one of the main challenges. In biometric systems, its most sensitive piece of information, the template, is constantly being exchanged between its building blocks. instead of having one template, in this paper we generate a set of synthetic templates to camouflage the genuine one. To test their indistinguishability, we suppose an attack and compare two different classifications results of reconstructed faces: humans and SVM classifier. For the former, we built a platform where testers could classify a set of random preimages reconstructed from real or synthetic (honey) templates. From an attacker point of view, we noticed that, compared to the SVM classifier, human testers showed better results in terms of classification distinguishability.