The authors leverage masked diffusion language models, which generate text by iteratively denoising blocks of tokens.
Guided diffusion enables trustworthy synthetic text with formal privacy guarantees
The authors show that steering the denoising process of diffusion language models with privatized semantic embeddings produces coherent synthetic text that preserves differential privacy and outperforms existing methods.
Big Tech
Qi Pang · Hanshen Xiao · Andre Manoel · Lipika Ramaswamy · Wenting Zheng · Virginia Smith · +1 more
NVIDIA · Carnegie Mellon University · Purdue University
Research Digest··3 min read
The authors propose Guided Diffusion, a framework that uses diffusion language models (DLMs) to generate synthetic text conditioned on semantic embeddings.
§