Euclidean-Distance Based Fuzzy Commitment Scheme for Biometric Template Security

With the introduction of triplet loss and other Euclidean-distance based loss functions, significant performance enhancement has been obtained for deep-learning based face recognition systems. However, existing template security solutions are based on binary feature vectors or binarization of real valued feature vectors using shielding function. This paper proposes a key-binding cryptographic template security scheme that uses lattice structure and sphere packing in the Euclidean space. In contrary to existing schemes, the proposed scheme can be applied to real-valued feature vectors. Therefore, it is more compatible with recent face recognition methods based on Euclidean distance. In this paper, two different versions of our proposed schemes are discussed in terms of security and complexity. Experimental investigations on Labeled Faces in the Wild dataset suggest no degradation in the performance of the face recognition system after being secured by our proposed scheme.