Reproduce “Differentially Private Fine-tuning of Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Da Yu, Saurabh Naik, Arturs Backurs, Sivakanth Gopi, Huseyin A. Inan, Gautam Kamath, Janardhan Kulkarni, Yin Tat Lee and 4 more

Researchers at Microsoft and affiliated universities developed simpler, sparser, and faster algorithms for differentially private fine-tuning of large language models. Their approach achieves state-of-the-art privacy-utility tradeoffs, often closely matching non-private model performance, and demonstrates that larger models unexpectedly maintain utility better under privacy constraints.

Key insights

  • A surprising finding is that larger models not only yield higher absolute utility but also incur a smaller relative utility drop when differentially private fine-tuning is applied, challenging previous assumptions about smaller models being more suitable for DP.
  • Restricting private updates to a low-dimensional manifold via PEFT methods effectively preserves the knowledge gained during public pre-training, mitigating the performance degradation typically associated with DP.
  • The modular nature of the fine-tuning framework allows a single large pre-trained model to be augmented with lightweight, privately-trained 'plug-in' parameter modules for various tasks, reducing storage and deployment complexities.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.