A Robust CycleGAN-L2 Defense Method For Speaker Recognition System

With the rapid development of voice technology, speaker recognition is becoming increasingly prevalent in our daily lives. However, with its increased usage, security issues have become more apparent. The adversarial attack poses a significant risk to model by making small changes input and thus causing neural network produce an incorrect output. Nevertheless, there are currently limited defense techniques for models. To this end, we propose robust CycleGAN-L2(CYC-L2) method. method automatically adjusts size dataset according learning generative networks on dataset, uses L2 loss functions constrain better faster training. In paper, will compare effectiveness against white-box attacks using existing defenses proposed. experimental results show that not only plays effect than other methods mentioned under x-vector but also does reduce accuracy benign examples closed-set identification.