Introducing TF-Encrypted

TF-Encrypted/TFE is an open-source framework for Secure Multi-party Computation (MPC) machine learning. The advantage of TFE is that it’s built on top of TensorFlow, allowing non-cryptographic experts to quickly experiment MPC machine learning, while leveraging all the advantages of TensorFlow’s optimizations, including graph compilation and distributed orchestration.

We have developed several projects based on TFE (e.g. The solution winning the iDASH2019 competition), and are actively contributing to TFE. In this blog we will describe a proof-of-concept use case, and give a walkthrough on how to do it with TFE.

MPC machine learning training in five minutes with TF-Encrypted

The use case: Collaborative fraud detection

Suppose Alice is a bank, Bob is a police office. Alice and Bob know many individuals in common, and both parties knows some information about the individuals from different aspects. E.g., Alice knows the individuals’ credit card bills, while Bob knows whether the individuals have some fraud history or not (denoted by label =1 or 0). Now Bob wants to build a fraud detection model with the help of Alice. Alice is willing to collaborate, but she consider her part of user information sensitive and is not willing to share them directly.

This problem could be summarized as a how to do secure collaborative machine learning training on a vertically split dataset. Here’s a simple walkthrough showing how it could be done efficiently using TFE.

Suppose the dataset contains 7000 samples with 16 features held by Alice, and the label held by Bob. A random generated example could be found here.

Step 1. Prepare three machines and set up environments

Check python3 and pip3 is correctly installed, then install TensorFlow and TFE on the three machines.

# python3 --version
Python 3.6.9
# pip3 --version
pip 9.0.1 from /usr/lib/python3/dist-packages (python 3.6)
#pip3 install tensorflow==1.13.2
#pip3 install tf-encrypted

Step 2. Edit the following file config.json

Replace machine:port with your own IP:Port. Make sure the three machines are able to access each other via IP:Port.

{
    "server0": "machine1:port1",
    "server1": "machine2:port2",
    "server2": "machine3:port3"
}

Step 3. Write TFE training code

We provide an example for Logistic Regression here.

Step 4. Copy the files to the same directory

Copy config.json , common.py , training_alice.py , aliceTrainFile.csv to machine1;

Copy config.json , training_bob.py , bobLabel.csv to machine2;

Copy config.json , training_server.py to machine3;

Step 5. Run!

Run the following command on the three machines, and the trained logistic regression model (the weights for each feature) will be printed on machine1.

python3 training_bob.py
python3 training_server.py
python3 training_alice.py

Extra notes

Using ABY3

The above example uses the default protocol of TFE, which is POND. If we want to use ABY3, we have to build from source:

pip3 uninstall tf-encrypted
git clone https://github.com/tf-encrypted/tf-encrypted.git
cd tf-encrypted
pip3 install -e .
make build   

Then change training_alice.py to this one here and re-run.

Production usage

TFE is an experimental software and must be hardened before used in production environments.