Cancelable Multibiometrics Template Security Using Deep Binarization and Secure Hashing

Abstract

Template security and privacy is of utmost significance while designing a biometric system. Several biometric template protection systems have been presented in the past, but none of them have succeeded in striking a compromise between matching performance and security. This paper proposes a hybrid template protection technique for a multibiometric system based on deep binarization and secure hashing. The technique is employed at different stages of multibiometric fusion. In particular, the proposed technique of multibiometric fusion for template protection is tested using face and electrocardiogram (ECG) biometrics. The pre-trained deep CNN model utilizes transfer learning to analyze both the biometrics and prepare multimodal templates at different stages of biometric fusion e.g. sensors, features, and matchers. The templates obtained from different states of fusion are mapped to their corresponding classes, which are represented as binary codes that are unique and randomly generated. The binary codes are further encrypted for noninvertibility using a cryptographic hash, and thus the information of fused templates is hidden. Finally, hash codes are used to perform matching. The evaluation of the proposed technique using database for face (Multi-PIE) and ECG (PTB) biometrics reports high accuracy satisfying the requirements of unlinkability, cancelability, and irreversibility for template protection.