MODIFIED BIOMETRIC TEMPLATE PROTECTION METHOD WITH NONLINEAR TRANSFORMATIONS

Abstract

Context. Biometric data is a common option for authentication or identification. However, it is vulnerable and not replaceable in case of stealing. Several methods for constructing protected biometric templates are proposed in literature, one of them is biohashing. However, linearity of biohashing may be a vulnerability. MLP-hash is similar, but adds nonlinearity. It is modified in this work.
Objective. The goal of this work is to develop a modification of MLP-hash which is faster and allows better separation of users by their templates.
Method. This work focuses on modifying MLP-hash, a biohashing variation with nonlinear transformations. One of the proposed changes is the usage of normalization before applying nonlinear transformation in each layer of MLP-hash. Different normalization methods are investigated and compared. The other proposed change is the simplification of the pseudorandom matrices used in each layer of MLP-hash. Each such matrix is replaced by a block matrix in which blocks that are laying on the diagonal are orthonormal matrices and all other blocks are filled with zeros. Each nonzero block is generated from the user’s secret token. In order to make the effect of each nonzero block less localized, a pseudorandom permutation is added before each matrix multiplication and also after all layers. Pseudorandom permutations are also generated with the user’s secret token as seed. The proposed method can be used in a similar way to how original MLP-hash and biohashing methods are used: it takes the user’s secret token and biometric vector of fixed length and outputs a binary vector of fixed length with the same or smaller dimensionality. MLP-hash with block matrices is compared to the original while applying different normalization techniques and different nonlinear transformations.
Results. The proposed modifications, original MLP-hash and biohashing have been implemented in code. Speed and accuracy of user separation with the usage of these methods have been compared on feature vectors extracted from fingerprints with the usage of Gabor filters.
Conclusions. The conducted experiments have shown an increase of speed and ability to separate user templates from the substitution of proposed block matrices and an increase of ability to separate user templates from the usage of normalization.
Comparison of different normalizations and nonlinear transformation has also been conducted. The practical usefulness of the developed method is that it is faster and can be used in applications when users expect no delays while still being difficult to invert. The prospects for further research include testing this method with other biometric modalities, other nonlinear transformations and normalization techniques and an analysis of inversion difficulty of the developed method in comparison to MLP-hash and biohashing.