Education

• PhD student at Carnegie Mellon University.
• M.S. in Shanghai Jiao Tong University, 2025
• B.S. in Shanghai Jiao Tong University, 2022

Research Interests

• Trustworthy AI: Mainly on copyright protection and authentication w./against AI personalization.
• Privacy and Security: Adversarial attacks and defenses, data extraction, and membership inference.
• Generative Modeling.

Research Experience

[ICML 2025] Data Extraction on Personalized Generative Models
Instructor: Steven Wu (CMU)
May. 2024 — Present

• Revealed the data leakage threat for large personalized generative models by performing successful exact data extraction, providing strong evidence of potential infringements.
• Utilized guidance between pretrained and fine-tuned models to generate within high-density regions of the fine-tuned data probability distribution, enabling precise data extraction.

[ICML 2023 (Oral) & CVPR 2024] Copyright Authentication and Imitation Prevention for Diffusion Models
Oct. 2022 — May. 2024
Instructor: Yang Hua (QUB), Hao Wang (LSU), Tao Song (SJTU)

• Crafted a visualizable copyright authentication mechanism by modeling and analyzing the conceptual difference between fine-tuned and pretrained diffusion models theoretically.
• Established a framework to craft adversarial watermarks against unauthorized diffusion model-based artwork mimicry by modeling adversarial attacks on diffusion models theoretically.

Selected Publications

Taming Outlier Tokens in Diffusion Transformers

Wu, X.*, Wang, Y.*, Fu, T.-J., et al. (2026). "Taming Outlier Tokens in Diffusion Transformers." arXiv:2605.05206.

Exploring Diffusion Models’ Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks

Wu, X.*, Zhang, J.*, Hua, Y., et al. (2026). "Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks." KDD 2026.

Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM

Wu, X., Pang, Y., Liu, T., and Wu, Z. S. (2025). "Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM." NeurIPS 2025.

Winning the MIDST Challenge: New Membership Inference Attacks on Diffusion Models for Tabular Data Synthesis

Wu, X., Pang, Y., Liu, T., and Wu, Z. S. (2025). "Winning the MIDST Challenge: New Membership Inference Attacks on Diffusion Models for Tabular Data Synthesis." TPDP 2025.

Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models

Wu, X., Zhang, J., and Wu, S. (2025). "Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models." ICML 2025.

Targeted Attack Improves Protection against Unauthorized Diffusion Customization

Zheng, B., Liang, C., and Wu, X. (2025). "Targeted Attack Improves Protection against Unauthorized Diffusion Customization." ICLR 2025. Spotlight presentation.

CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion

Wu, X., Hua, Y., Liang, C., et al. (2024). "CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion." CVPR 2024.

Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples

Liang, C.*, Wu, X.*, Hua, Y., et al. (2023). "Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples." ICML 2023. Oral presentation.

Mist: Towards Improved Adversarial Examples for Diffusion Models

Liang, C.* and Wu, X.* (2023). "Mist: Towards Improved Adversarial Examples for Diffusion Models." arXiv:2305.12683.

Service

• Reviewers for NeurIPS 2024, ICLR 2025.