In the ASVspoof2021 physical access (PA) task, due to the mismatch between the simulated training data and the evaluation data from the real scenario, performance of previous top-performing countermeasure systems had a significant degradation. The main reason for this phenomenon can be attributed to a simulation-to-real gap. In this work, the effect of sim-to-real gap is investigated on different datasets for replay attacks. Differences in the frequency domain between simulated and real datasets are investigated to cross the sim-to-real gap. On the basis of our previous work, different sub-band acoustic features have different capabilities in distinguishing spoof utterances from bonafide ones. To decrease the effect of sim-to-real gap and build a robust anti-spoofing system against the replay attacks, a cross-subband countermeasure is proposed in this work. Furthermore, we use visualized heatmap to explore the artefacts captured by model trained with cross-subband method. To verify the generalization capability of the cross-subband method on different datasets, several sets of comparative experiments were also done. The results show that our cross-subband countermeasure is robust to sim-to-real gap in the PA task, and the fusion model based on it is regarded as one of the top-performing antispoofing systems in the ASVspoof2021 Challenge.
Cite as: Lu, J., Zhang, Y., Wang, W., Zhang, P. (2022) Robust Cross-SubBand Countermeasure Against Replay Attacks. Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 126-132, doi: 10.21437/Odyssey.2022-18
@inproceedings{lu22_odyssey,
author={Jingze Lu and Yuxiang Zhang and Wenchao Wang and Pengyuan Zhang},
title={{Robust Cross-SubBand Countermeasure Against Replay Attacks}},
year=2022,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2022)},
pages={126--132},
doi={10.21437/Odyssey.2022-18}
}