Research map / Federated learning and privacy / Differential privacy
Privacy Preserving In-Context-Learning Framework for Large Language Models
Bishnu Bhusal, Manoj Acharya, Ramneet Kaur, Colin Samplawski, Anirban Roy, Adam D. Cobb, Rohit Chadha, Susmit Jha
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use proposed clip-average-blend DP aggregation instead of Amin et al. 2024 | private synthetic ICL demos · TREC 4-shot ICL (eps=1) | accuracy (%): 39.2 -> 63.0 | Improved |
| use proposed aggregation instead of Amin et al. 2024 | private synthetic ICL demos · AGNews 4-shot ICL (eps=inf) | accuracy (%): 76.8 -> 73.5 | Hurt |
| use proposed aggregation instead of Amin et al. 2024 | private synthetic ICL demos · MIT-D 4-shot ICL (eps=1) | accuracy (%): 84.5 -> 79.9 | Hurt |
| use proposed aggregation instead of Amin et al. 2024 | private synthetic ICL demos · DBPedia 4-shot ICL (eps=1) | accuracy (%): 75.1 -> 76.8 | No clear effect |
| blend private logits with public-prompt logits | private synthetic ICL demos · several ICL tasks | accuracy (%) | Improved |
| increase privacy budget eps (1 to inf) (ablation) | private synthetic ICL demos · DBPedia ICL (4-shot) | accuracy (%): 76.80 -> 80.94 | Improved |
| increase privacy budget eps (1 to inf) on TREC (ablation) | private synthetic ICL demos · TREC ICL (1-shot) | accuracy (%): 68.32 -> 62.40 | Hurt |
| increase LLM size/quality (Llama-3.2-1B to 8B) (ablation) | private synthetic ICL demos · DBPedia ICL (eps=1, model sweep) | accuracy (%): 53.9 -> 89.9 | Improved |
| use DP synthetic demonstrations instead of real private samples | private synthetic ICL demos · Enron membership inference | MIA AUC: 94.20 -> 51.78 | Improved |
| use proposed aggregation instead of Amin et al. 2024 for JSON | private synthetic generation · WikiMoviesJSON (eps=1) | validates (%): 74.2 -> 81.1 | Improved |
Closest papers by meaning
- Privacy-Preserving In-Context Learning for Large Language Models (ICLR 2024 · similarity 0.93)
- Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation (ICLR 2024 · similarity 0.92)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.92)
- SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.91)
- Data-adaptive Differentially Private Prompt Synthesis for In-Context Learning (ICLR 2025 · similarity 0.90)
- InvisibleInk: High-Utility and Low-Cost Text Generation with Differential Privacy (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
