EPTLENet: Replay Attack Detection with Efficient Parameter Transfer Learning Based on ERes2Net

Existing approaches for speech replay attack detection still face the problem of poor robustness. This is due to the small scale of the dataset and the lack of generalizability to unseen spoofing devices. In response to the issues above, this paper proposes an Efficient Parameter Transfer Learning method based on ERes2Net, named EPTLENet, to enhance the accuracy and generalization of replay attack detection. Firstly, the robustness of the algorithm is enhanced by utilizing the ERes2Net model. ERes2Net primarily modifies the Res2Net blocks to strengthen the feature fusion mechanism. Specifically, features within individual residual blocks are integrated to enhance local information interaction. Additionally, acoustic features from different stages are aggregated as inputs to incorporate global information, and the receptive field of EPTLENet is increased through Attention Feature Fusion (AFF) in this paper. Secondly, the problem of small dataset scale is overcome by transferring pre-trained weights using deep transfer learning methods, thereby enhancing the accuracy of the algorithm. Experimental results demonstrate that ERes2Net outperforms Res2Net and ResNet_SE on the ASVspoof 2017_2.0 dataset to detect replay attacks. Furthermore, combining the above three models with Deep Transfer Learning (DTL) can further improve accuracy. The proposed EPTLENet outperformed other systems by achieving a 36% reduction in Equal Error Rate (EER) compared to the baseline system.