Bicentric One-Class Learning With Directed Attention Fusion for Generalizable Speech Spoofing Detection

Abstract

Generalizable speech spoofing detection requires capturing complex variations in temporal and spectral domains, but their simple fusion strategies limit the joint modeling of underlying physical and statistical characteristics artifacts. Meanwhile, current deep learning-based binary classification methods rely heavily on prior knowledge of spoofing attacks, which biases the learning process toward seen samples and significantly degrades generalization to unseen attacks. To this end, we propose a bicentric one-class learning with directed attention fusion network (BicOC-DAFNet). BicOC-DAFNet introduces directed attention fusion (DAF) into the fusion layer. DAF dynamically aggregates different contextual information from diverse receptive fields during spectral-temporal fusion. Simultaneously, bicentric one-class softmax (BicOC-softmax) supervises learning of a discriminative feature space: genuine representations aggregate into a compact cluster around a dedicated centroid, while spoofing representations diverge bidirectionally opposite to a spoofing centroid and away from the one-class orientation, ensuring spoofing space coverage while preventing overfitting. We further validate the module’s adaptability across both end-to-end and pretrained front-ends. Experimental results demonstrate that our method outperforms current methods, while comprehensive evaluations on 9 authoritative test datasets yield a pooled equal error rate (EER) of 3.90%.