Piranha: A GPU Platform for Secure Computation

Piranha is a C++-based platform for accelerating secure multi-party computation (MPC) protocols on the GPU in a protocol-independent manner. It is designed both for MPC developers, providing a modular structure for easily adding new protocol implementations, and secure application developers, allowing execution on any Piranha-implemented protocols. This repo currently includes a secure ML inference and training application, which you can find in /nn.

Piranha is described in more detail in our USENIX Security '22 paper! If you have questions, please create git issues; for eventual replies, you can also reach out to jlw@berkeley.edu.

Warning: This is an academic proof-of-concept prototype and has not received careful code review. This implementation is NOT ready for production use.

Build

This project requires an NVIDIA GPU, and assumes you have your GPU drivers and the NVIDIA CUDA Toolkit already installed. The following has been tested on AWS with the Deep Learning Base AMI (Ubuntu 18.04 ) Version 53.5 AMI.

1. Checkout external modules

git submodule update --init --recursive

1. Build CUTLASS

cd ext/cutlass
mkdir build
cmake .. -DCUTLASS_NVCC_ARCHS=<YOUR_GPU_ARCH_HERE> -DCMAKE_CUDA_COMPILER_WORKS=1 -DCMAKE_CUDA_COMPILER=<YOUR NVCC PATH HERE>
make -j

1. Install GTest. We use it for unit testing.

sudo apt install libgtest-dev libssl-dev
cd /usr/src/gtest
sudo mkdir build
cd build
sudo cmake ..
sudo make
sudo make install

2. Create some necessary directories

mkdir output; mkdir files/MNIST; mkdir files/CIFAR10

3. Download the MNIST/CIFAR10 datasets, if using. This step might take a while

cd scripts
sudo pip install torch torchvision
python download_{mnist, cifar10}.py

4. Build Piranha at a specific fixed point precision and for a particular protocol. 3-party replicated secret sharing is the default and doesn't require a command-line flag.

make -j8 PIRANHA_FLAGS="-DFLOAT_PRECISION=<NBITS> -D{TWOPC,FOURPC}"

Run

1. Copy and set up a run configuration using config.json as a base. It is already set up to perform a 10-epoch SecureML training run; simply specify party IPs in the configuration.

2. Run Piranha on each machine with a party number (0 -> n_parties - 1):

./piranha -p <PARTY NUM> -c <CONFIG FILE>

Running locally

You may want to run Piranha on a local machine for development. An example configuration for 3-party local execution can be found at files/samples/localhost_config.json with an accompanying runfile. You can modify the runfile to change which GPUs Piranha uses for each party using the CUDA_VISIBLE_DEVICES environment variable. The script uses GPUs 0-2 by default, but can be changed to run on a single GPU as well. Note that due to contention, hosting several parties on a single GPU will limit the problem sizes you can test and incur some additional overhead.

Start the computation with:

./files/samples/localhost_runner.sh

Citation

You can cite the paper using the following BibTeX entry (the paper links to this repo):

@inproceedings {watson22piranha,
    author = {Watson, Jean-Luc and Wagh, Sameer and Popa, Raluca Ada},
    title = {Piranha: A {GPU} Platform for Secure Computation},
    booktitle = {31st USENIX Security Symposium (USENIX Security 22)},
    year = {2022},
    isbn = {978-1-939133-31-1},
    address = {Boston, MA},
    pages = {827--844},
    url = {https://www.usenix.org/conference/usenixsecurity22/presentation/watson},
    publisher = {USENIX Association},
    month = aug,
}

Artifact Evaluation

For our experiments, we use a cluser of AWS GPU-provisioned machines. Reviewers should have credentials to access the environment, but due to resource limits, we can only support one reviewer evaluating at a time. You can run Piranha to regenerate Figures 4, 5, 6, and 7, as well as Tables 2, 3, and 4.

Evaluation runs through experiments/run_experiment.py, which should be executed on the control instance we provide with the required dependencies. Here are the relevant options:

usage: run_experiment.py [-h] [--start] [--stop] [--figure FIGURE] [--table TABLE] [--generate] [--fast] [--verbose]

Run artifact evaluation!

optional arguments:
  -h, --help       show this help message and exit
  --start          Provision cluster for experiments. _Please suspend the cluster while not running experiments :)_
  --stop           Suspend evaluation machines.
  --figure FIGURE  Figure # to run.
  --table TABLE    Table # to run.
  --generate       Generate figure/table images.
  --fast           Run all the (relatively) fast runs, see README for more information
  --verbose        Display verbose run commands, helpful for debugging

• You can start and stop the cluster with --start and --stop, respectively. Please use these if you're not running evaluation! GPU instances are not cheap and cost about $450/day to keep running.

• Use the --figure and --table flags to run data generation for each of the paper's figures/tables. They're fairly automatic and should run without intervention.

• Generate each figure/table with the --generate flag. You can run the evaluation script on partial results and the results will reflect those partial values. Figures generate .png files in artifact_figures/artifact while table replication generates JSON. You can compare to the paper figures/tables generated into artifact_figures/paper from hardcoded data.

• Very important note on timing. Unfortunately, MPC still requires a significant amount of time (~30 hrs/training run) on a larger network like VGG16. A conservative estimate is that for Figure 5 alone, > 270 computation-hours are required to replicate the full figure. We've included a --fast flag if you'd like to replicate every other datapoint first (will still require a number of compute-hours), then come back to the VGG-based values.

• Use --verbose if something isn't working and you want to take a look at the raw output or need an error message. In the backend, we use Ansible to communicate with each of the machines in the cluster.