Extracting Discriminative Binary Template for Face Template Protection

This paper addresses the security issues of the face bio-metric templates stored in a database. In order to improve the security level of the stored face templates, cryptographic techniques are commonly employed. Since most of such techniques require a binary input, thresholding is usually employed to binarize the real valued face features. While binary templates are obtained, the discriminability of the original templates may be affected and so is the recognition performance. In order to overcome this limitation, this pa-per proposes a new approach to discriminative extract bi-nary templates from original face templates. A projection is applied before thresholding such that the thresholding can fit the projected data distribution better, which makes the extracted binary templates more discriminative. A dis-criminability index is constructed to measure the discrim-inability of the extracted binary templates and the projec-tion matrix is optimized. The proposed method is evaluated on three public domain databases, namely FERET, CMU-PIE and FRGC. Experimental results show that, in compar-ison to the existing thresholding-based methods, the pro-posed method improves the GAR by around 11 % − 13 % at a FAR of 1%.