Membership Inference Attack Against Music Diffusion Models via Generative Manifold Perturbation

Yuxuan Liu, Peihong Zhang, Rui Sang, Zhixin Li, Yizhou Tan, Yiqiang Cai, Shengchen Li · 2026

Read Membership Inference Attack Against Music Diffusion Models via Generative Manifold Perturbation by Yuxuan Liu, Peihong Zhang, Rui Sang, Zhixin Li, Yizhou Tan, Yiqiang Cai, Shengchen Li (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-01645.txt.

Description

Yuxuan Liu and colleagues audit generative music models for copyright compliance by testing whether an audio clip trained the model. Loss based signals like reconstruction error line up poorly with perception, giving weak separation at the low false positive rates forensics demands. Their white box LSA-Probe measures a geometric trait of reverse diffusion: the smallest time normalized perturbation budget that pushes an intermediate state across a perceptual degradation threshold. Training members sit in more stable regions and demand a higher degradation cost, making that cost a membership signal.

Catalog record

CREATORYuxuan Liu, Peihong Zhang, Rui Sang, Zhixin Li, Yizhou Tan, Yiqiang Cai, Shengchen Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22071
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.01645

Browse the full library · About the Tower of Babel