Research on adversarial sample generation methods for speaker recognition
Aiming at the problem that adversarial samples generated by generative adversarial sample generation methods have low authenticity and poor attack effect,an adversarial sample generation method ACGAN based on AdvGAN and CGAN is proposed.First,attacking a specific target,ACGAN generates targeted adversarial samples in the frequency domain by introducing additional target labels in the training and attack stages.Secondly,the gated convolutional neural network is introduced in the generator and discriminator to help the ACGAN model capture more accurate data features,thereby improving the success rate of the attack.Finally,the perceptual loss function is introduced to minimize the difference in speech feature representation between the model output and the target output,thereby improving the auditory quality of the generated samples.Experimental results show that compared with the existing methods in targeted attacks,the ASR is improved by 1.5%,and the SNR and PESQ are improved by 10.5%and 11.1%respectively,which proves the effectiveness and potential of ACGAN in the field of adversarial sample generation.
adversarial examplesgeneratordiscriminatorgated convolutional neural networkperceptual loss
万方数据
维普
