ORPHEUS: A Separation-Robust Proactive Defense for Singing Voice Conversion

Zhaolin Wei, School of Cyber Science and Engineering, Wuhan University; Dengpan Ye, School of Cyber Science and Engineering, Wuhan University, and Cyberspace Institute of Advanced Technology, Guangzhou University; Yanjiao Chen, College of Electrical Engineering, Zhejiang University; Jiacheng Deng, Ziyi Liu, Yuhan Lin, and Yunna Lv, School of Cyber Science and Engineering, Wuhan University; Yueyun Shang, School of Mathematics and Information Science, Guangzhou University; Zhihong Tian, Cyberspace Institute of Advanced Technology, Guangzhou University

Recent advances in singing voice conversion enable realistic cloning of a singer's voice, raising concerns about unauthorized voice misuse. Existing proactive defenses inject imperceptible perturbations to disrupt such conversion, and they are designed for vocal signals rather than mixed music. In real-world scenarios, however, attackers usually obtain mixed music and apply source separation to decompose the mixture into vocal and backing tracks before conversion. Since separation models are trained to distinguish vocals from background components, perturbations are often treated as backing, causing the perturbations to be filtered out. To address this, we present ORPHEUS, the first proactive defense framework tailored for singing voice conversion involving source separation. Specifically, we jointly (i) interfere with the separation model via mask-misguiding and cross-track losses, (ii) disrupt identity information using an ensemble of heterogeneous speaker encoders to enhance transferability, and (iii) optimize perceptual quality through tonality harmony constraints and psychoacoustic masking. To evaluate ORPHEUS, we conduct experiments on two datasets with three source separation models and four singing voice conversion systems. Compared with the best-performing baseline, ORPHEUS further reduces identity similarity by 8.02% and 5.61% on two speaker verification models, respectively, demonstrating consistently stronger defense effectiveness and cross-model transferability.

USENIX Security '26 Open Access Sponsored by
King Abdullah University of Science and Technology (KAUST)

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.

BibTeX
@inproceedings {320769,
author = {Zhaolin Wei and Dengpan Ye and Yanjiao Chen and Jiacheng Deng and Ziyi Liu and Yuhan Lin and Yunna Lv and Yueyun Shang and Zhihong Tian},
title = {{ORPHEUS}: A {Separation-Robust} Proactive Defense for Singing Voice Conversion},
booktitle = {35th USENIX Security Symposium (USENIX Security 26)},
year = {2026},
isbn = {978-1-939133-58-8},
address = {Baltimore, MD},
pages = {4345--4363},
url = {https://www.usenix.org/conference/usenixsecurity26/presentation/wei-zhaolin},
publisher = {USENIX Association},
month = aug
}