A preliminary study of factors affecting the performance of a Playback Attack Detector

A playback attack refers to the situation when an intruder plays back a recording of a true client uttering his or her pass phrase in order to bypass a speaker verification system. Playback attacks are effective and easy to implement, thus, posing a serious threat to the security of a pass-phrase protected speaker verification system. A Playback Attack Detector (PAD) has been proposed to safeguard the speaker verification systems against playback attacks. The proposed PAD capitalizes on the hypothesis that due to the random nature of speech production process, distinctions can always be found between any two utterances. Thus, if little distinction is found between two utterances, it is likely that one utterance is the playback of the other. In this paper, the proposed PAD algorithm is summarized. The design and the collection process of a database used to evaluate the effects of different factors (e.g. client, pass phrase, playback speaker, communication channel, and number of stored recordings) on the PAD's performance is detailed. Experimental procedures used for performance evaluation are described and results are presented.