Reproduce “Privacy Preserving In-Context-Learning Framework 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

Bishnu Bhusal, Manoj Acharya, Ramneet Kaur, Colin Samplawski, Anirban Roy, Adam D. Cobb, Rohit Chadha, Susmit Jha

Researchers at the University of Missouri and SRI International developed a privacy-preserving framework for large language models that generates high-quality synthetic text for in-context learning. This method achieves competitive accuracy on text classification and information extraction tasks under strong differential privacy guarantees, and empirically resists PII extraction and membership inference attacks.

Key insights

  • A novel logit aggregation and blending approach, which clips individual logits and averages them with public logits, provides strong DP guarantees without relying on complex, threshold-based mechanisms like the Sparse Vector Technique (SVT).
  • Leveraging advanced composition and KV-cache reuse for fixed, disjoint data subsets enables efficient, linear-time generation of coherent and semantically meaningful synthetic text sequences, optimizing privacy budget allocation.
  • Competitive utility in various ICL tasks (e.g., text classification, structured data generation) can be achieved under strict privacy budgets (e.g., $\epsilon=1$), demonstrating an improved privacy-utility trade-off compared to prior methods.

Summary by alphaXiv.

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