B-CAT: Breaking the Barrier Between Communication Rounds and Accuracy for Private Transformer Inference

B-CAT: Breaking the Barrier Between Communication Rounds and Accuracy for Private Transformer Inference This repository implements the paper: B-CAT: Breaking the Barrier Between Communication Rounds and Accuracy for Private Transformer Inference based on CrypTen and SHAFT. Installing B-CAT The following commands run successfully on Ubuntu 22.04 with Python 3.10.12.We recommend using conda to manage your Python environment(Our work is based on the shaft.environment). 0. Set up Conda Environment (Recommended) conda create -n b-cat python=3.10.12 -yconda activate b-cat 1. Install Dependencies pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118pip install wheel==0.40.0 2. Install B-CAT Open the webpage Anonymous Repository and download the source code zip, then unzip. cd B-CAT-Anonymous pip install . 3. Install Transformers (for Hugging Face Integration) git clone -b 'v4.45.0' --depth 1 https://github.com/huggingface/transformerspip install ./transformers Running Experiments We have a set of sub-directories in the examples directory for reproducible experimental results. Additional dependencies for the experiments are included in the requirements.txt file in each subdirectory under the folder. Please refer to the README.md file in the sub-directories for instructions on how to set up and run the experiments. 1. unit-test - Costs of private all of our protocols. 2. op_plain_acc_test - Accuracy comparison test of operators of B-CAT and other comparision of schemes under plaintext. Since the README in the uploaded file cannot be modified, please refer to this version. License B-CAT is MIT licensed, as found in the LICENSE file.