上海科技大学知识管理系统

Voice communication services, such as VoIP and teleconferences, have emerged as a popular tool for frauds. Also development of audio deepfakes has heightened concerns about the authenticity and integrity of the content of voice calls. However, anonymity of VoIP calls and  make it difficult to protect network audio streams using existing audio authentication techniques. Thus in this paper, we propose an audio authentication system that can enable live speakers to protect their speeches from malicious alterations. The core idea is directly embedding content-based signatures in audio streams or recordings. This system adopts deep learning and audio compression techniques to generate audio signatures, which can resist signal distortion brought by network transmission and audio compression. We evaluate it against modern audio forgery techniques, and the result shows that our designs could protect the integrity of live audio streams and recordings.