Biometric Template Protection Using Deep Learning

Abstract

Use of biometrics in digital society has raised the questions of biometric template protection and secure authentication. The biometric template protection mechanisms known so far hardly maintain a trade-off between security of template database and recognition performance. This paper proposes a hybrid technique of template protection for a multibiometric system that provides better matching performance and infallible from fraudulent attacks. The multimodal system is prepared from face and ECG biometrics. The ECG as a biometrics not only supplements the face biometrics in a multimodal system but also ensures security for robust recognition. The pre-trained deep learning models are used to process both biometrics and prepare multimodal templates. The templates are mapped to their corresponding classes represented by randomly generated unique binary codes. These binary codes are further encrypted using cryptographic hash for non-invertiblity and hide information of fused templates. Finally, the matching is performed using hash codes for ensuring an additional layer of defense against adversarial attacks.