Improving replay attack detection by combination of spatial and spectral features

Abstract

In this paper, we propose a replay attack detection based on score fusion of spatial and spectral features-based systems. Recently, a replay attack detection (RAD) system using generalized cross-correlation (GCC) of a stereo signal has been proposed. The GCC is calculated from non-speech sections of input signals. It reported that the GCC-based method achieved high performance under several situations. However, since the performance of the GCC-based method depends on the situations, it is required to improve the performance without situation dependence. The GCC-based method uses spatial features, which utilize the different feature from spectral features. In this paper, we perform score fusion of the GCC-based and the spectral feature-based methods to improve the robustness of RAD systems. In the experiments, the proposed method achieved a relative error reduction of 69.5%, compared with a GCC-based single method under one of the hard tasks. And, the performance of score fusion systems improved without situation dependence.