Reproduce “Differentially Private Subspace Fine-Tuning for Large 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
Lele Zheng, Xiang Wang, Tao Zhang, Yang Cao, Ke Cheng, Yulong Shen
DP-SFT is a two-stage framework for differentially private fine-tuning of Large Language Models, which projects gradient updates into a low-dimensional subspace before noise injection. This approach yields up to 11.51% higher accuracy over prior methods on NLP benchmarks under standard privacy budgets, significantly improving utility while maintaining strong privacy guarantees.
Key insights
- The effective updates during LLM fine-tuning primarily reside within a compact, low-dimensional subspace of the gradient space.
- Confining differentially private noise injection to this identified low-dimensional subspace significantly mitigates the 'curse of dimensionality' and reduces noise magnitude.
- These task-specific subspaces exhibit transferability across related tasks, allowing for their construction using public datasets without consuming the private training budget.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.