InferDPT: Privacy-Preserving Inference for Closed-box Large Language Model¶
Authors: Meng Tong; Kejiang Chen; Jie Zhang; Yuang Qi; Weiming Zhang; Nenghai Yu; Tianwei Zhang; Zhikun Zhang
Year: 2023
Venue: arXiv 2023
type: system collection: catalogue status: auto
Links: arXiv:2310.12214
Summary¶
Auto-extracted summary (unreviewed)
InferDPT is a framework for privacy-preserving inference with black-box (closed) LLMs such as GPT-4 that applies differential privacy to the prompt text: a perturbation module replaces prompt tokens with a DP mechanism (RANTEXT, built on a random adjacency list over the GPT-4 token vocabulary) and a local extraction module running a small open LLM (Vicuna/Llama) recovers coherent text from the perturbed remote generation. On CNN/Daily Mail, WikiText-103 and an arXiv dataset, generation quality is comparable to non-private GPT-4, and RANTEXT dominates SANTEXT+ and CUSTEXT+ on the privacy/utility trade-off, keeping over 0.90 protection against embedding inversion attacks at epsilon = 6.0.
Abstract¶
Large language models (LLMs), like ChatGPT, have greatly simplified text generation tasks. However, they have also raised concerns about privacy risks such as data leakage and unauthorized data collection. Existing solutions for privacy-preserving inference face practical challenges related to computation time and communication costs. In this paper, we propose InferDPT, the first practical framework for the privacy-preserving Inference of black-box LLMs, implementing Differential Privacy in Text generation. InferDPT comprises two key modules: the "perturbation module" utilizes the exponential mechanism to generate a perturbed prompt, facilitating privacy-preserving inference with black-box LLMs, and the "extraction module", inspired by knowledge distillation and retrieval-augmented generation, extracts coherent and consistent text from the perturbed generation result, ensuring successful text generation completion. To address privacy concerns related to previous exponential mechanisms' susceptibility to embedding revision attacks, we introduce RANTEXT, a novel differential privacy mechanism integrated into the perturbation module of InferDPT, which introduces the concept of "RANdom adjacency" for TEXT perturbation within the prompt. Experimental results across three datasets demonstrate that the text generation quality of InferDPT is comparable to that of non-private GPT-4, and RANTEXT surpasses existing state-of-the-art mechanisms, namely, SANTEXT+ and CUSTEXT+ in the trade-off between privacy and utility. Even with an privacy parameter epsilon value of 6.0, RANTEXT achieves an average privacy protection rate exceeding 90% against embedding revision attacks, which is 0.58 times higher than that of SANTEXT+ and 3.35 times higher than that of CUSTEXT+.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | InferDPT framework for privacy-preserving inference with black-box LLMs via differential privacy on prompts | llm | 0.85 |
| contribution | RANTEXT differentially private text perturbation mechanism based on random adjacency lists | llm | 0.85 |
| contribution | Extraction module using a local open-source LLM to produce coherent output from perturbed generations | llm | 0.85 |
| contribution | Evaluation against BERT inference, embedding inversion and an adaptive GPT-4 inference attack | llm | 0.85 |
| schemes | differential privacy | llm | 0.85 |
| models | GPT-4 | llm | 0.85 |
| models | Vicuna-7B | llm | 0.85 |
| models | LLaMA-2-7B | llm | 0.85 |
| models | LLaMA-3.1-8B | llm | 0.85 |
| models | BERT | llm | 0.85 |
| models | text-embedding-ada-002 | llm | 0.85 |
| datasets | CNN/Daily Mail | llm | 0.85 |
| datasets | WikiText-103 | llm | 0.85 |
| datasets | arXiv dataset | llm | 0.85 |
| tasks | text generation | llm | 0.85 |
| tasks | private LLM inference / prompt privacy | llm | 0.85 |
| hardware | NVIDIA RTX A6000 GPU | llm | 0.85 |
| hardware | Intel Xeon Gold 6130 CPU | llm | 0.85 |
| security level | differential privacy, headline epsilon = 6.0 (epsilon 0.01-18 evaluated) | llm | 0.85 |
| threat model | untrusted black-box LLM service provider; adversary tries to reconstruct raw prompt tokens | llm | 0.85 |
| interaction | other | llm | 0.85 |
| code url | https://github.com/mengtong0110/InferDPT.Index (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/SaitejaUtpala/dp (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/openai/tiktoken (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/mengtong0110/InferDPT | llm | 0.85 |
| llm relevance | high | llm | 0.85 |
| key result | privacy protection vs embedding inversion attack: > 0.90 average (RANTEXT, epsilon = 6.0; 0.58x higher than SANTEXT+, 3.35x higher than CUSTEXT+) | llm | 0.85 |
| key result | privacy protection vs BERT inference attack: > 80% (RANTEXT, epsilon from 0.01 to 18.0) | llm | 0.85 |
| key result | privacy protection vs BERT inference attack (relative): 1.11x SANTEXT+, 1.41x CUSTEXT+ (epsilon = 18.0, CNN/Daily Mail) | llm | 0.85 |
| key result | text generation quality: comparable to non-private GPT-4 (three datasets) | llm | 0.85 |
| key result | 1-gram privacy leakage rate in perturbed generation: 0.081 (RANTEXT epsilon = 6.0 (SANTEXT+ 0.103, CUSTEXT+ 0.099, non-private 0.110)) | llm | 0.85 |
System card: InferDPT¶
| Group | 6. Trusted hardware, statistical and integrity-only alternatives |
| Family | Text-level DP perturbation |
| Arithmetic | Native |
| Interaction | Non-interactive |
| Model scale | GPT-4 (black box) |
| Status | Active |
| Relation to the SoK | Works with closed APIs without server cooperation. |
| Note | Utility drops at strong privacy levels. No cryptographic protection. |
| Systems page | InferDPT |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: Differential privacy 44 [p.1], Leveled / somewhat HE 13 [p.2], FINAL 2 [p.12]
- model: GPT 49 [p.1], BERT 13 [p.9], CNN 12 [p.5], LLaMA 2 [p.3], Transformer 1 [p.13]
- hardware: NVIDIA RTX 1 [p.9], Intel Xeon 1 [p.9], GPU 1 [p.9]
- technique: Knowledge distillation 2 [p.2], Fine-tuning 2 [p.13], Two-party computation 2 [p.13]
- task: Language modeling / generation 42 [p.1], Medical / healthcare 8 [p.12], Finance 6 [p.14]
- topic: Large language models 65 [p.1], Private / secure inference 9 [p.1], Open source 7 [p.1], Homomorphic encryption 2 [p.13]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/6-087_inferdpt-privacy-preserving-inference-for-closed-box-large-l.pdf |
catalogue | 17 | 1.7 MB | 6c72a38ea9cc |
Database key: inferdpt-privacy-preserving-inference-for-closed-box-large-language · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set inferdpt-privacy-preserving-inference-for-closed-box-large-language COLUMN VALUE