Securing Face Templates using Deep Convolutional Neural Network and Random Projection

Abstract

Rapidly growing use of biometrics across enterprises, consumer electronics, and applications has made biometric template protection indispensable to any practical biometric system. To provide biometric template protection, several methods have been reported but most of them have a trade-off between matching performance and the template security. To address this problem, we improvise upon existing methods to provide a method for face template protection with improved matching performance and high template security. The proposed method uses deep Convolutional Neural Network (CNN) together with random projection to maximize the inter-user variations, reduce the dimensionality (thus eliminating redundancy) of the extracted feature vector of each face image and minimize the intra-user variations. The proposed method is robust enough to perform even with one-shot enrollment of users. Three publicly available face datasets, namely, CMU-PIE, FEI and Color-FERET are used for evaluation. The proposed method achieves ~99.5% Genuine Accept Rate (GAR) at zero False Accept Rate (FAR), improving the matching performance by ~10% over the comparable state-of-the-art methods, while providing high template security.