Attacks on Speech Biometric Authentication

We investigate the security of DTW, VQ and GMM methods that have been used in speaker authentication systems. We present attack models based on adversary knowledge. We start with naive adversaries without knowledge of an authentic speaker and develop them into highly knowledgeable adversaries who know the speaker’s information, have the speaker’s voice samples, acquire the speaker’s template, and know an algorithm of the speaker authentication system. We propose an analysis-synthesis forgery in which the informed adversary can exploit information such as feature vectors from the template and a statistical probability from the voice samples of the target speakers to re-generate a forgery that can be used in remote or on-line authentication. We evaluate our scheme with two speech datasets. The results show that the proposed scheme outperforms other attack models reported in the literature.