Re-assessing the threat of replay spoofing attacks against automatic speaker verification

This paper re-examines the threat of spoofing or presentation attacks in the context of automatic speaker verification (ASV). While voice conversion and speech synthesis attacks present a serious threat, and have accordingly received a great deal of attention in the recent literature, they can only be implemented with a high level of technical know-how. In contrast, the implementation of replay attacks require no specific expertise nor any sophisticated equipment and thus they arguably present a greater risk. The comparative threat of each attack is re-examined in this paper against six different ASV systems including a state-of-the-art iVector-PLDA system. Despite the lack of attention in the literature, experiments show that low-effort replay attacks provoke higher levels of false acceptance than comparatively higher-effort spoofing attacks such as voice conversion and speech synthesis. Results therefore show the need to refocus research effort and to develop countermeasures against replay attacks in future work.

I. Introduction

Spoofing refers to the presentation of a falsified or manipulated sample to the sensor of a biometric system in order to provoke a high score and thus illegitimate acceptance. In recent years, the automatic speaker verification (ASV) community has started to investigate spoofing and countermeasures actively [1], [2]. A growing body of independent work has now demonstrated the vulnerability of ASV systems to spoofing through impersonation [3], voice conversion [4], [5], speech synthesis [6], [7] and attacks with non-speech, artificial tone-like signals [8].