Despite tremendous progress in speaker verification recently, replay spoofing attacks are still a major threat to these systems. Focusing on dataset-specific scenarios, anti-spoofing systems have achieved promising in-domain performance at the cost of poor generalization towards unseen out-of-domain datasets. This is treated as a domain mismatch problem with a domain adversarial training (DAT) framework, which has previously been applied to enhance generalization. However, since only one domain discriminator is adopted, DAT suffers from the false alignment of cross-domain spoofed and genuine pairs, thus failing to acquire a strong spoofing-discriminative capability. In this work, we propose the dual-adversarial domain adaptation (DADA) framework to enable fine-grained alignment of spoofed and genuine data separately by using two domain discriminators, which effectively alleviates the above problem and further improves spoofing detection performance. Experiments on the ASVspoof 2017 V.2 dataset and the physical access portion of BTAS 2016 dataset demonstrate that the proposed DADA framework significantly outperforms the baseline model and DAT framework in cross-domain evaluation scenarios. It is shown that the newly proposed DADA architecture is more robust and effective for generalized replay attack detection.
Cite as: Wang, H., Dinkel, H., Wang, S., Qian, Y., Yu, K. (2020) Dual-Adversarial Domain Adaptation for Generalized Replay Attack Detection. Proc. Interspeech 2020, 1086-1090, doi: 10.21437/Interspeech.2020-1255
@inproceedings{wang20l_interspeech,
author={Hongji Wang and Heinrich Dinkel and Shuai Wang and Yanmin Qian and Kai Yu},
title={{Dual-Adversarial Domain Adaptation for Generalized Replay Attack Detection}},
year=2020,
booktitle={Proc. Interspeech 2020},
pages={1086--1090},
doi={10.21437/Interspeech.2020-1255}
}