Replay attack: Its effect on GMM-UBM based text-independent speaker verification system

Abstract

Spoofing is a technique to bias the system decisions towards target speakers without the physical presence of genuines. The focus of spoofing is on target speakers, resulting a sharp increase in false acceptance. The availability of high quality recording and playback devices (i.e. smart phones), have made spoof attacks by replay data a most common and easy approach, even without much knowledge about speech processing. In this work we analyze and demonstrate the seriousness of replay attacks, and then examine the vulnerability of classic GMMUBM based speaker verification (SV) system to replay data by experimental studies. Since no standard database is available, we develop a replay dataset of 34 speakers from Indian Institute of Technology Guwahati - Multi-Variability (IITG-MV) speech database, that covers all kind of variabilities. To avoid any gender biasing, we made our studies for male and female speakers independently and finally in together. Under replay attacks, for genuine 459 male and 220 female trials, the equal error rate (EER) increases from 2.61% to 15.03% and 1.82% to 34.09%, respectively. In total, the EER increases from 2.50% to 20.76%. As obvious, the degradation in EER is mainly due to increase in false acceptance ratio (FAR). The case is worsen for female speakers. It may be due to the fine spectral structures of female speakers. These observations reveal that GMM-UBM based SV system is highly vulnerable to replay attacks and needs wider attentions.