Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models

FineXtract approximates fine-tuning as a gradual shift from a pretrained diffusion model toward the fine-tuning distribution. Extrapolating along that shift guides generation toward high-probability regions of the private data, while clustering identifies likely training examples.

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