GHOST

Official repo for the paper "Mitigating Gradient Inversion Risks in Language Models via Token Obfuscation" in Asia CCS'26.

Usage

Install Dependencies

Create a new conda environment and install the required packages. The following command will create a new conda environment named GHOST with Python 3.12.3 and install the necessary packages.

First, make sure you have conda installed. You can download and install Miniconda from their official websites.

Then, run the following commands to create a new conda environment and activate it.

conda create -n GHOST python=3.12.3
conda activate GHOST

Then, you need to install torch for most of the deep learning operations. Our expected version is 2.3.0 with cuda 12.1. However, the compatibility depends on your specific GPU hardware. You need to check the official PyTorch website and see if you can install these versions. If not, try a different version and it may still work. But this is out of our testing scope.

pip install torch==2.3.0 --index-url https://download.pytorch.org/whl/cu121

Then, you need to pip install the following packages.

pip install transformers==4.44.2 datasets==2.19.1 evaluate==0.4.3 accelerate==0.30.1 nltk==3.8.1 spacy==3.8.2 absl-py==2.1.0 rouge_score==0.1.2 scikit-learn==1.6.0 bitsandbytes==0.42.0 peft==0.14.0 vec2text==0.0.13 flair==0.15.1

After that, you need to download the en_core_web_sm model for spacy. You can do this by running the following command:

python -m spacy download en_core_web_sm

You also need to download the punkt tokenizer for nltk. You can do this by running the following command:

python -c "import nltk; nltk.download('punkt')"

Experiments