Voice Spoofing Detection using Feature Adaptive One-Class Classification Based on Graph Neural Network
Haiyan Lan, Yike Li, Yuhua Wang, L. Y. Dong, Zhiying Han, Kai Yang · 2024
To address the problems of complex real voice feature distribution, easy overfitting by learning only one classification boundary, and poor generalization ability of existing voice spoofing detection methods for unknown attacks, this paper combines the representation ability of graph neural networks with the learning objective of one-class classification, and proposes a feature adaptive one-class voice detection framework based on graph neural networks. The core idea of the feature adaptive one-class classification loss (FAOC) in the framework is to cluster the real voices affected by similar factors more closely, learn several classification boundaries of the real voices, and optimize the distance of the real voices with different distributions, so that the model can obtain more general features of the real voices. At the same time, the relationship between different frequency bands or time segments of the voice is modeled by using graph neural networks, and the graph embedding features with strong recognition ability are obtained, which can better exploit the performance advantages of FAOC. A series of experiments are conducted on the ASVspoof2019 LA and ASVspoof2021 LA datasets, and the results on the ASVspoof2021 LA show that the EER and min-tDCF are improved by 45.8% and 48.2% compared to the baseline. The results show that the proposed method has good performance for voice synthesis attack detection.