Spoofing detection in the wild: an investigation of approaches to improve generalisation

The generalisation of spoofing detection solutions to spoofing attacks or recording conditions not seen in training data has been a focus since the inception of research in this area. We report our investigation of three strategies to improve upon generalisation, namely data augmentation, the fine-tuning of a pre-trained model, and a Siamese model with a cross-attention mechanism. When evaluated under domain-mismatched conditions, we show that these techniques are all effective in reducing model overfitting and in encouraging the learning of more generalisable models by capturing the (di)similarity between bonafide or spoofed test and known-to-be bonafide reference utterances. Evaluations using the in-the-wild dataset show that our model achieves a relative improvement of almost 60% compared to the best results reported in the literature.