Practical Adversarial Attacks Against Speaker Recognition Systems

Unlike other biometric-based user identification methods (e.g., fingerprint and iris), speaker recognition systems can identify individuals relying on their unique voice biometrics without requiring users to be physically present. Therefore, speaker recognition systems have been becoming increasingly popular recently in various domains, such as remote access control, banking services and criminal investigation. In this paper, we study the vulnerability of this kind of systems by launching a practical and systematic adversarial attack against X-vector, the state-of-the-art deep neural network (DNN) based speaker recognition system. In particular, by adding a well-crafted inconspicuous noise to the original audio, our attack can fool the speaker recognition system to make false predictions and even force the audio to be recognized as any adversary-desired speaker. Moreover, our attack integrates the estimated room impulse response (RIR) into the adversarial example training process toward practical audio adversarial examples which could remain effective while being played over the air in the physical world. Extensive experiment using a public dataset of 109 speakers shows the effectiveness of our attack with a high attack success rate for both digital attack (98%) and practical over-the-air attack (50%).