My Music My Choice: Adversarial Protection Against Vocal Cloning in Songs

Ilke Demir, Gerald Pena Vargas, Alicia Unterreiner, David Ponce, Umur A. Ciftci

Keywords: adversarial protection; audio defense; voice cloning; vocal replication; adversarial generation

TL;DR: MMMC protects vocal tracks in songs by learning to generate imperceptible adversarially protected versions, which in turn breaks voice cloning methods, leading to very low quality outputs.

Abstract: Vocal cloning poses unprecedented threats to the music industry, enabling unauthorized reproduction of artists' vocal characteristics in songs. We introduce My Music My Choice (MMMC), a lightweight adversarial protection framework designed to proactively defend against vocal cloning in musical contexts. MMMC generates imperceptibly modified audio that preserves original vocals while significantly degrading the output of generative AI. MMMC achieves high-quality protected vocals with 0.944 STOI, reducing cloned output quality to 0.558 STOI. Our approach demonstrates effective transfer across different musical styles and maintains protection when vocals are reconstructed into full songs, providing a practical defense mechanism for artists and music creators.

Track: Paper Track

Confirmation: Paper Track: I confirm that I have followed the formatting guideline and anonymized my submission.

Submission Number: 11