VOLERE: Leakage Resilient User Authentication Based on Personal Voice Challenges

Voiceprint Authentication as a Service (VAaS) offers great convenience due to ubiquity, generality, and usability. Despite its attractiveness, it suffers from user voiceprint leakage over the air or at the cloud, which intrudes user voice privacy and retards its wide adoption. The literature still lacks an effective solution on this issue. Traditional methods based on cryptography are too complex to be practically deployed while other approaches distort user voiceprints, which hinders accurate user identification. In this paper, we propose a leakage resilient user authentication cloud service with privacy preservation based on random personal voice challenges, named VOLERE (VOice LEakage REsilient). It applies a novel voiceprint synthesis method based on a Log Magnitude Approximate (LMA) vocal tract model to fuse original voices of different speaking modes in order to generate a synthesized voiceprint for authentication. Thus, raw voiceprints of users can be well protected. We implement VOLERE and conduct a series of user tests. Experimental results show sound performance of VOLERE regarding authentication accuracy, efficiency, stability, leakage resilience and user acceptance. In particular, its authentication accuracy is reasonably stable regardless user nationality, gender, age, elapsed time, and environment, as well as variance of speaking modes.