GForce

GForce: GPU-Friendly Oblivious and Rapid Neural Network Inference

A crypto-assisted framework leveraging GPU, additive secret share, and homomorphic encryption to protect the privacy of models and queries in inference.

The details of this project are presented in the following paper:

GForce: GPU-Friendly Oblivious and Rapid Neural Network Inference
Lucien K. L. Ng, Sherman S. M. Chow
Usenix Security Symposium 2021

Suggested Setup

We tested our code on 2 Google Cloud VMs located in the same region. They were running with

• Ubuntu 18.04 LTS
• Nvidia V100 GPU
• 8 Virtual Intel Xeon (Skylake) CPUs at 2GHz
• 52 GB RAM

Note that GPU is necessary. Nvidia P100 GPU is also well suited with our code.

How to Install

1. Install Anaconda, a Python package manager.

2. Build an Anaconda virtual environment and install required packages as follows:

   conda env create -f conda_requirement.yml
   

   It will create a virtual environment called gforce. Please activate this environment by:

   conda activate gforce
   

3. Install SEAL-Python by following the instructions in its README.md.

   Remember to activate the virtual environment gforce (by conda activate gforce) before installation.

How to run

Testing the Performance of Neural Networks

• Run VGG-16 on CIFAR-10

  • If you want to run both the server-side and client-side code in a local machine, run the following command

    python src/secure_vgg.py
    

  • If the server and the client are on different machines, please run:

    • The server run the following command, where $IpClient is the IP address of the client

      python src/secure_vgg.py --ip="$IpClient" -s 0
      

    • The client run the following command, where $IpServer is the IP address of the server

      python src/secure_vgg.py --ip="$IpServer" -s 1
      

• If you want to try running other neural networks (e.g., MiniONN's NN with MaxPool/AvgPool) or testing on CIFAR-100, you may add the argument --test $TestName. $TestName can be:

  • vgg_cifar10
  • vgg_cifar100
  • minionn_avgpool (for CIFAR-10)
  • minionn_maxpool (for CIFAR-10)

  For example, the command of running VGG-16 on CIFAR-100 on a single machine:

  python src/secure_vgg.py --test vgg_cifar100
  

  You may combine with the arguments --ip="$IpClient" -s 0 and --ip="$IpServer" -s 1 to test the cross-machine performance.

Testing the Performance of Non-Linear Layers

You may also try testing the performance of non-linear layers.

• For ReLU Layer:

  python src/relu_dgk.py
  

• For Maxpool Layer (with 2x2 pooling windows):

  python src/maxpool2x2_dgk.py
  

The arguments python --ip="$IpClient" -s 0 and --ip="$IpServer" -s 1 for testing the cross-machine performance.

Testing the Accuracy of Neural Networks

For testing the accuracy of our pre-trained GForce DNN (under modulo operations), you may run

python src/swalp_inference.py --test $TestName

where $TestName can be

• vgg16_cifar10
• vgg16_cifar100
• minionn_avgpool (for CIFAR-10)
• minionn_maxpool (for CIFAR-10)

Disclaimer

DO NOT USE THIS SOFTWARE TO SECURE ANY REAL-WORLD DATA OR COMPUTATION!

This software is a proof-of-concept meant for performance testing of the GForce framework ONLY. It is full of security vulnerabilities that facilitate testing, debugging, and performance measurements. In any real-world deployment, these vulnerabilities can be easily exploited to leak all user inputs.

Related Project

Our team also proposed Goten: GPU-Outsourcing Trusted Execution of Neural Network Training, a secure solution for privacy-preserving neural network training.

Acknowledgement

We also thanks Anna P. Y. Woo for her countless effort in overcoming the issues for integrating GForce with SWALP, which resulted in impressive model accuracy in CIFAR-10/100 and generated the experimental data in the paper.