Spoofing Speaker Verification System by Adversarial Examples Leveraging the Generalized Speaker Difference

Speaker verification system has gained great popularity in recent years, especially with the development of deep neural networks and Internet Things. However, security speaker based on not been well investigated. In this paper, we propose an attack to spoof state-of-the-art generalized end-to-end (GE2E) loss function for misclassifying illegal users into authentic user. Specifically, design a novel deploy generator generating effective adversarial examples slight perturbation then these achieve our goals. The success rate can reach 82% when cosine similarity is adopted deep-learning-based system. Beyond that, experiments also reported signal-to-noise ratio at 76 dB, which proves that higher imperceptibility than previous works. summary, results show only neural-network-based but more importantly ability hide from human hearing or machine discrimination.