Research map / Federated learning and privacy / Differential privacy
Flocks of Stochastic Parrots: Differentially Private Prompt Learning for Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 one-shot prompt data in a prompt (membership inference attack) | GPT3-Babbage prompted LLM · dbpedia | MIA AUC: 0.84 | Hurt |
| private soft prompt instead of non-private soft prompt (eps=8 vs inf) | DP Soft-Prompt · sst2 | accuracy: 95.64 -> 92.31 | Hurt |
| private soft prompt instead of non-private soft prompt (eps=8 vs inf) | DP Soft-Prompt · mnli | accuracy: 82.49 -> 75.15 | Hurt |
| prefix instead of soft prompt under eps=8 | DP Prefix · mnli | accuracy: 75.15 -> 80.57 | Helped |
| prefix instead of soft prompt under eps=8 | DP Prefix · sst2 | accuracy: 92.31 -> 91.97 | No clear effect |
| DP prompt tuning versus DP full fine-tuning on a simple task | DP Soft-Prompt · sst2 | accuracy: 85.89 -> 92.31 | Helped |
| DP prompt tuning versus DP full fine-tuning on a hard task | DP Soft-Prompt · mnli | accuracy: 83.30 -> 75.15 | Hurt |
| RoBERTa-Large instead of RoBERTa-Base for DP prefix | DP Prefix · mnli | accuracy: 80.57 -> 86.40 | Helped |
| use zero-shot vs PromptPATE with IID public data | PromptPATE · dbpedia | test accuracy: 44.2 -> 80.3 | Helped |
| OOD public data (agnews) instead of IID public data (ablation) | PromptPATE · dbpedia | test accuracy: 80.3 -> 74.6 | Hurt |
| OOD public data (imdb) instead of IID public data (ablation) | PromptPATE · sst2 | test accuracy: 88.8 -> 87.2 | No clear effect |
| use GPT3-Curie instead of GPT3-Babbage | PromptPATE · sst2 | test accuracy: 87.2 -> 92.7 | Helped |
Closest papers by meaning
- DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.90)
- Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs (ICML 2025 · similarity 0.88)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.88)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.88)
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · similarity 0.87)
- DPZero: Private Fine-Tuning of Language Models without Backpropagation (ICML 2024 · similarity 0.87)
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.