Cancellable template generation for speaker recognition based on spectrogram patch selection and deep convolutional neural networks

Nowadays, biometric systems have replaced password-or token-based authentication systems in many fields to improve the security level. However, biometric systems are also vulnerable to security threats. Unlike passwords, biometric templates cannot be replaced if lost or compromised. To deal with issue of compromising biometric templates, template protection schemes have evolved to make it possible to replace the biometric templates. A cancellable biometric scheme is such a template protection scheme that can replace a biometric template, when it is stolen or lost. The biometric used here is speech. It is important to preserve user confidentiality. Cancellable biometrics is a new notion addressed for this problem. This paper presents a scheme for cancellable speaker recognition based on spectrogram patch selection. The simulation results reveal that the suggested approach is practical, and it satisfies the desired criteria such as renewability, security and performance. The accuracy of the proposed cancellable speaker recognition scheme reaches 98.75% with a Convolutional Neural Network (CNN) composed of three layers.