Random permutation Maxout transform for cancellable facial template protection

Abstract

In this paper, we propose a salting based two-factor cancelable biometrics construct, dubbed Random Permutation Maxout (RPM) transform for facial template protection. The RPM transform is inspired from a member of rank-based Locality Sensitive Hashing (LSH), namely Winner Takes All hashing, which was devised for data retrieval. With externally generated user-specific parameters, RPM converts a continuous facial feature vector into a max ranked indices vector as cancellable template. Since the features magnitude of facial features have been transformed to the discrete index form, the resulting template is robust against noises and it is strongly concealed from the adversary learning on the original facial features. This lays a strong promise on non-invertibility requirement The LSH theory compliance RPM is shown minimal performance deterioration after transform. The experimental results render reasonable accuracy performance on benchmark AR and FERET datasets. We also perform several rigorous security, privacy, revocability and unlinkability analyses, which are required for cancellable biometrics techniques.