Replay attack detection based on deformable convolutional neural network and temporal-frequency attention model

Abstract As an important identity authentication method, speaker verification (SV) has been widely used in many domains, e.g., mobile financials. At the same time, existing SV systems are insecure under replay spoofing attacks. Toward a more secure and stable system, this article proposes attack detection system based on deformable convolutional neural networks (DCNNs) time–frequency double-channel attention model. In DCNN, positions of elements kernel not fixed. Instead, they modified by some trainable variable to help model extract useful local information from input spectrograms. Meanwhile, domino is adopted effective distinctive features collect valuable for distinguishing genuine speeches. Experimental results ASVspoof 2019 dataset show that proposed can detect attacks accurately.