An Experimental Analysis on Mapping Strategies for Cepstral Coefficients Multi-projection in Voice Spoofing Detection Problem

Abstract

With the cheapening of certain electronic devices, the use of biometric user authentication systems is becoming increasingly common. In particular, the use of microphones for user recognition is a reality for many people in the most common tasks, such as accessing their own cell phone or even their own bank account at an ATM. However, recent works demonstrate that the practice of frauds such as the improper presentation of a recording of an authentic user to the recognition system can pose a threat to the security of these systems. As a countermeasure to such practices, liveness detection techniques have emerged with the intention of ensuring that the signal presented to the system comes from the human vocal tract or from a recording. Techniques based on the analysis of spectrograms, which consist of a matrix representation of the relationship between the frequency of the signal and its temporal duration, are widely used in this topic. However, little study has been carried out on the use of mapping operators on such representations in order to make feature vectors belonging to $\mathbb {R}^n$ and, consequently, facilitate its definition and use in most binary classifiers. In this work, we perform an experimental analysis on different mapping operators applied in the construction of handcrafted features on spectrograms calculated by various techniques. In addition, we also analyze the effect of fusing and projecting such features. Finally, we were able to observe that, with the use of the proposed material, which consists of a tool of simple definition and low complexity, we obtained competitive results to those presented by the baselines of the theme.