Skip to content

Split-and-Denoise: Protect Large Language Model Inference with Local Differential Privacy

Authors: Peihua Mai; Ran Yan; Zhe Huang; Youjia Yang; Yan Pang
Year: 2024
Venue: ICML 2024

type: system collection: catalogue status: auto

Links: arXiv:2310.09130

Summary

Auto-extracted summary (unreviewed)

Split-N-Denoise (SnD) protects embedding-as-a-service LLM inference by running the token-embedding layer on the client, perturbing the token embeddings with d-chi (local differential privacy) noise before sending them to the server, and then denoising the returned output embeddings on the client with a pre-trained denoise model, without modifying the LLM parameters. Across BERT, GPT-2 and T5 model families on CoLA, QQP, MRPC and RTE, SnD improves downstream utility under the same privacy budget by over 10% on average compared with token-embedding privatization, text-to-text privatization and RAPT baselines.

Abstract

Large Language Models (LLMs) excel in natural language understanding by capturing hidden semantics in vector space. This process enriches the value of text embeddings for various downstream tasks, thereby fostering the Embedding-as-a-Service (EaaS) business model. However, the risk of privacy leakage due to direct text transmission to servers remains a critical concern. To address this, we introduce Split-N-Denoise (SnD), an private inference framework that splits the model to execute the token embedding layer on the client side at minimal computational cost. This allows the client to introduce noise prior to transmitting the embeddings to the server, and subsequently receive and denoise the perturbed output embeddings for downstream tasks. Our approach is designed for the inference stage of LLMs and requires no modifications to the model parameters. Extensive experiments demonstrate SnD's effectiveness in optimizing the privacy-utility tradeoff across various LLM architectures and diverse downstream tasks. The results reveal an improvement in performance under the same privacy budget compared to the baselines by over 10\% on average, offering clients a privacy-preserving solution for local privacy protection.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution split inference framework that runs the token embedding layer client-side llm 0.85
contribution local DP (d-chi privacy) perturbation of token embeddings before transmission llm 0.85
contribution client-side denoising model trained on raw token representations and specific noises llm 0.85
contribution no modification of LLM parameters required llm 0.85
contribution evaluation across BERT, GPT-2 and T5 families with mutual-information privacy analysis llm 0.85
schemes local differential privacy (d-chi privacy) llm 0.85
schemes split inference llm 0.85
models DistilBERT llm 0.85
models BERT-base llm 0.85
models BERT-large llm 0.85
models T5-small llm 0.85
models T5-base llm 0.85
models T5-large llm 0.85
models GPT-2 small llm 0.85
models GPT-2 medium llm 0.85
models GPT-2 large llm 0.85
models GPT-2 XL llm 0.85
datasets CoLA llm 0.85
datasets QQP llm 0.85
datasets MRPC llm 0.85
datasets RTE llm 0.85
datasets TweetEval Offensive llm 0.85
datasets Hate Speech 18 llm 0.85
datasets Health Fact llm 0.85
datasets Daily Dialogue llm 0.85
tasks private LLM embedding-as-a-service inference llm 0.85
tasks sentence classification llm 0.85
tasks paraphrase / pair similarity llm 0.85
tasks textual entailment llm 0.85
threat model client-server split inference; server is untrusted and sees only LDP-perturbed token embeddings llm 0.85
interaction other llm 0.85
code url https://github.com/NusIoraPrivacy/eaas-privacy.Our (URL found in the full text) regex 0.50
code url https://github.com/PolyAI-LDN/task-specific-datasets.Chatzikokolakis (URL found in the full text) regex 0.50
code url https://github.com/NusIoraPrivacy/eaas-privacy llm 0.85
llm relevance high llm 0.85
key result utility improvement over baselines: > 10% on average (same privacy budget, across LLM architectures and downstream tasks) llm 0.85
key result accuracy (CoLA): 0.688 at eta=100 vs 0.751 unperturbed (BERT-base (110M), SnD) llm 0.85
key result accuracy (RTE): 0.592 at eta=100 vs 0.616 unperturbed (BERT-base (110M), SnD) llm 0.85
key result accuracy (CoLA): 0.693 at eta=100 vs 0.766 unperturbed (GPT-2 XL (1.5B), SnD) llm 0.85
key result accuracy (QQP): 0.69 at eta=1 vs 0.72 unperturbed (T5-base (220M), SnD) llm 0.85

System card: Split-N-Denoise (SnD)

Group 6. Trusted hardware, statistical and integrity-only alternatives
Family Local DP on embeddings + client denoiser
Arithmetic Native
Interaction Non-interactive (single round)
Model scale Llama-2, GPT-2
Status Active
Relation to the SoK Formal DP but only statistical privacy.
Note Orders of magnitude cheaper than CKKS, but leaks information within the privacy budget.
Systems page Split-N-Denoise (SnD)

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: Differential privacy 46 [p.1], MPC 2 [p.5]
  • model: GPT 43 [p.2], BERT 14 [p.6], k-NN 7 [p.7], Transformer 3 [p.2], LLaMA 1 [p.9]
  • dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 47 [p.6], AG News 4 [p.14], SQuAD 3 [p.12], Yelp 3 [p.15], SST-2 1 [p.9]
  • hardware: TPU 8 [p.16], GPU 7 [p.15], NVIDIA V100 5 [p.16], NVIDIA RTX 2 [p.15], Intel Xeon 2 [p.15], CUDA 2 [p.15]
  • technique: Fine-tuning 22 [p.2], Split learning / split inference 13 [p.1], Federated learning 3 [p.2], Attention 2 [p.15]
  • task: Sentiment analysis 4 [p.14], Finance 4 [p.14], Medical / healthcare 3 [p.11], Text classification 2 [p.12]
  • topic: Large language models 60 [p.1], Benchmark 6 [p.5], Homomorphic encryption 3 [p.1], Open source 3 [p.2], Private / secure inference 2 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/6-086_split-and-denoise-protect-large-language-model-inference-wit.pdf catalogue 22 1.4 MB bb1af83c8a79

Database key: split-and-denoise-protect-large-language-model-inference-with-local · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set split-and-denoise-protect-large-language-model-inference-with-local COLUMN VALUE