Yuxuan Liu

PhD Candidate in AI Security & Music Information Retrieval

Xi'an Jiaotong-Liverpool University

Research Projects

Interactive demonstrations of cutting-edge research in AI security and music analysis

ISMIR 2025

MAIA: Music Adversarial Inpainting Attack

Y. Liu, P. Zhang, R. Sang, Z. Li, S. Li

Importance-driven adversarial attacks via music inpainting for MIR systems. Uses GACELA inpainting model with adversarial guidance for targeted modifications.

92.8% ASR MOS: 4.0/5 8 Audio Samples
View Interactive Demo →
ICASSP (Under Review)

LSA-Probe: Latent Stability Adversarial Probe

Y. Liu, P. Zhang, R. Sang, Z. Li, Y. Tan, Y. Cai, S. Li

Membership inference for music diffusion models via generative manifold perturbation. Measures adversarial cost across diffusion timesteps for robust detection.

TPR@1%FPR: +8% AUC: 0.67 Interactive Timestep Slider
Explore Algorithm →
AAAI Workshop (Accepted)

TS-RaMIA: Token-Structure-Aware Membership Inference

Y. Liu, P. Zhang, R. Sang, Z. Li, S. Li

Structure-aware membership inference attacks for music generation models. Exploits token-level and structural patterns for improved detection.

32.7% TPR@5%FPR 10 Demo Samples ROC Visualizations
Training-time Protection

Music Unlearnable for MusicLDM

Y. Liu

Adversarial perturbations that prevent diffusion models from learning your music style. Protects music through training-time CLAP embedding misalignment.

CLAP-based Interactive Demo Training-time Defense
Explore Method →