Revealing privacy risks in music generation through transcription structure analysis
Scroll to explore
Music generation models learn from thousands of compositions. But can we tell if a specific piece was in the training set?
Why This Matters:
Hover over samples to see if they're training data
A three-step attack leveraging structural tokens in music notation
Extract structure-revealing tokens from ABC notation (bars, time signatures, key signatures)
Focus on the hardest-to-predict tokens (Top-64) which reveal memorization
Combine scores across different temperatures (0.8, 1.0, 1.2, 1.5) for robust detection
Final Score = Fusion(ScoreT=0.8, ScoreT=1.0, ScoreT=1.2, ScoreT=1.5)
See the attack in action on real music samples
Our method significantly outperforms baseline approaches
Detect if copyrighted music was used in model training
Verify compliance with data usage agreements
Evaluate privacy risks before deployment
@article{ts-ramia2025,
title={TS-RaMIA: Membership Inference Attacks
for Symbolic Music Generation Models},
author={Yuxuan Liu},
year={2025}
}