Revealing the Unseen: Guiding Personalized Diffusion Models to Expose Training Data

This paper introduces FineXtract, a framework for extracting training data from fine-tuned diffusion models shared online. By approximating the model’s distribution shift during fine-tuning and using clustering techniques, FineXtract can extract data with high quality, demonstrating potential risks of data leakage and copyright infringement.

Recommended citation: Wu X, Zhang J, Wu S. (2024). "Revealing the Unseen: Guiding Personalized Diffusion Models to Expose Training Data." arXiv preprint arXiv:2410.03039.
Download Paper