Anti Leakage: Protecting Privacy Hidden in Our Speech

In recent years, smart terminals have spread to every corner of our life. While smart terminals bring us much more convenience, they also take the risk of privacy-leakage because they may record our speech and analyze more informaiton. This paper proposes an emotion-protecting mechanism which is cross-model and available in reality to protect users’ speech privacy. Specifically, when devices(such as Alexa) receive uers’ voice command, the initial emotion would be converted to another emotion by our mechanism. As a result, the attackers can not infer the correct emotion from stolen speech data. To realize this mechanism, we design an algorithm which can convert original emotion into target emotion while keep the emotion distribution stable before and after filtering.We further design an unique training strategy to maximize our privacy protection effectiveness, which can improve the performance of our mechanism. Finally, comprehensive experiments conducted on online tests demonstrate the effectiveness of our mechanism with outstanding performance.