Skip to content

[ICLR'24 Spotlight] DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer

License

Notifications You must be signed in to change notification settings

VITA-Group/DP-OPT

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer

License: MIT

Official PyTorch Code for Paper: "DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer" Junyuan Hong, Jiachen T. Wang, Chenhui Zhang, Zhangheng Li, Bo Li, Zhangyang Wang, ICLR (Spotlight, top-5%) 2024.

paper / code / blog

TL;DR: We proposed the first end-to-end privacy-preserving automatic prompt engineering method.

Overview

featured

Large Language Models (LLMs) have emerged as dominant tools for various tasks, particularly when tailored for a specific target by prompt tuning. Nevertheless, concerns surrounding data privacy present obstacles due to the tuned prompts' dependency on sensitive private information. A practical solution is to host a local LLM and optimize a soft prompt privately using data. Yet, hosting a local model becomes problematic when model ownership is protected. Alternative methods, like sending data to the model’s provider for training, intensify these privacy issues facing an untrusted provider. In this paper, we present a novel solution called Differentially-Private Offsite Prompt Tuning (DP-OPT) to address this challenge. Our approach involves tuning a discrete prompt on the client side and then applying it to the desired cloud models. We demonstrate that prompts suggested by LLMs themselves can be transferred without compromising performance significantly. To ensure that the prompts do not leak private information, we introduce the first private prompt generation mechanism, by a differentially-private (DP) ensemble of in-context learning with private demonstrations. With DP-OPT, generating privacy-preserving prompts by Vicuna-7b can yield competitive performance compared to non-private in-context learning on GPT3.5 or local private prompt tuning.

Get Started

Prepare conda env.

conda create --name dp-opt python=3.8 -y
conda activate dp-opt
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install transformers datasets accelerate sentencepiece scikit-learn wandb autodp
# transformers==4.28.1

Prepare DLN datasets

bash setup_data.sh

To use openai models, create openai_config.py in the root folder. This will be only used for evaluation.

import openai

openai.api_key = "<your-key>"
# openai.organization = "<your-org>"
openai.api_base = "https://api.openai.com/v1"
openai_model_types = ['text-davinci-003']

⚠️ Warning: Setting echo and logprobs simultaneously is no longer supported for certain OpenAI models. However, classification inference with openai models requires both settings. Consider to host your own models, e.g., thru vLLM, instead.

Example: Do prompt engineer on website:

pip install gradio
python web_demo.py
# open http://127.0.0.1:7860
Train Test

Example: Use local model (lmsys/vicuna-7b-v1.3) to generate a instruction and test the instruction by OpenAI model (text-davinci-003).

  • OPT:
# generate a instruction
python train_opt.py --ape_mode=iid_ibwd --ensemble_gen=True --gen_temp=1.1 --num_prompt=40 --max_new_tokens=50 \
--data=sst2 --holdout_ratio=0.01
# evaluate the instruction
python eval_opt.py --ape_mode=iid_ibwd --ensemble_gen=True --gen_temp=1.1 --num_prompt=40 --max_new_tokens=50 \
--data=sst2 \
--test_model=text-davinci-003
  • DP-OPT:
# generate a instruction
python train_opt.py --ape_mode=iid_ibwd --ensemble_gen=True --gen_temp=1.1 --num_prompt=40 --max_new_tokens=50 \
--data=sst2 --holdout_ratio=0.01 \
--target_eps=8. --dp_eps=1.8 --dp_delta=5e-7 --tokenwise_gen=True
# evaluate the instruction
python eval_opt.py --ape_mode=iid_ibwd --ensemble_gen=True --gen_temp=1.1 --num_prompt=40 --max_new_tokens=50 \
--data=sst2 \
--target_eps=8. --dp_eps=1.8 --dp_delta=5e-7 --tokenwise_gen=True \
--test_model=text-davinci-003

Experiments

Wandb sweeps files are under sweeps/<data_name>/<method>.yml. sweeps/<data_name>/<method>.yml is used for tuning prompts. We use sweeps/<data_name>/<method>_test.yml to test prompts on different models.

Supported datasets: sst2, trec, mpqa, disaster.

image

Methods (exmaplified on sst2):

  • 5-shot In-Context Learning (ICL)
wandb sweep sweeps/sst2/icl.yml
  • Deep Language Network with One-layer (DLN-1)
wandb sweep sweeps/sst2/dln1.yml
wandb sweep sweeps/sst2/dln1_test.yml
  • Offsite Prompt Tuning (OPT)
wandb sweep sweeps/sst2/opt.yml
wandb sweep sweeps/sst2/opt_test.yml
  • Differentially-Private Offsite Prompt Tuning (DP-OPT)
wandb sweep sweeps/sst2/dp-opt.yml
wandb sweep sweeps/sst2/dp-opt_test.yml

Part of the codes are based on deep-language-networks.