Face Template Protection Using Deep Convolutional Neural Network

The growing use of biometrics has led to rising concerns about the security and privacy of the biometric data (template) since it is unique to each individual and cannot be replaced. To address this problem, many biometric template protection algorithms have been reported but most have a trade-off between matching performance and template security. In this work, we propose a method for face template protection, which improves upon existing face template protection algorithm, to provide better matching performance. The proposed method uses deep Convolutional Neural Network (CNN), with one-shot and multi-shot enrollment, to learn a robust mapping from face images of the users to the unique binary codes (assigned to the users during enrollment phase). The cryptographic hash (like SHA-3 512) of the user's binary code represents the protected face template. The deep CNN is trained to minimize the intra-class variations and maximize the inter-class variations. During verification, given an input face image of a user, deep CNN predicts the binary code assigned to the user. The hash of the predicted binary code is matched with the hash of the actual binary code assigned to the user during enrollment. Three face datasets, namely CMU-PIE, FEI and Color FERET are used for evaluation. The proposed method improves the matching performance by ~6% and reduces Equal Error Rate by about 4 times when compared to related work, while providing high template security.