Can We Trust Your Voice? Exploring Vulnerabilities in Voice Authentication

As voice authentication technology becomes more prevalent, its security flaws and vulnerabilities are garnering increasing scrutiny. State-of-the-art deep neural network (DNN) systems for voice authentication can achieve an accuracy of over 95%. However, DNN-based models are known to be vulnerable to attacks such as adversarial examples and data poisoning. An adversary may also take advantage of the limited generalization of current DNN-based models to circumvent the system, only requiring authentic voices to impersonate others. In this paper, we leverage a data poisoning attack and two voice authentication models to investigate the vulnerability and corresponding impacts on individual user and system security. We introduce a new toolkit, the Voice Authentication Poisoning Impact Evaluator (VAPIE), incorporating conventional machine learning and VGG-based models. VAPIE is designed to predict the potential impacts of various data poisoning scenarios launched by different attack-ers and to evaluate overall system security, achieving an accuracy rate of over 70 %. This facilitates a deeper understanding and mitigation of the risks associated with voice authentication technologies.