Learning to Fool the Speaker Recognition

Due to the widespread deployment of fingerprint/face/speaker recognition systems, risk in these especially adversarial attack, has drawn increasing attention recent years. Previous researches mainly studied attack vision-based such as fingerprint and face recognition. While for speech-based systems not been well yet, although it widely used our daily life. In this article, we attempt fool state-of-the-art speaker model present attacker , a lightweight multi-layer convolutional neural network well-trained by adding imperceptible perturbations onto raw speech waveform. We find that system is vulnerable achieve high success rate on both non-targeted targeted attack. Besides, an effective method leveraging pretrained phoneme optimize obtain tradeoff between perceptual quality. Experimental results TIMIT LibriSpeech datasets demonstrate effectiveness efficiency proposed model. And experiments frequency analysis indicate high-frequency more than low-frequency which different from conclusion previous image-based works. Additionally, ablation study gives insights into