Reading In The Dark: Classifying Encrypted Digits with Functional Encryption

This code implements the cryptographic part of Partially Encrypted Machine Learning using Functional Encryption (to appear at NIPS 2019). The code for the Collateral Learning component can be found here.

Requirements

Install PBC and Charm

Follow this tutorial.

Set up a database

Follow the beginning of this tutorial to create a database and an account and update the core/discretelogarithm file. Default is user dbuser with password password and database discretelogarithm.

Hashing

Discrete log precomputation assumes a specific environment:

export PYTHONHASHSEED=0

Tinker

If you want to be able to generate the confusion matrix, you will need tinker. On Ubuntu:

sudo apt-get install python3-tk

Python requirements

Installing the required python packages should be as easy as:

pip3 install -r requirements.txt

Generate your keys and fill you database

cd mnist
python3 initialization.py

This might take a while. Note that the progress bar that will appear often does not progress linearly: the first few percentage points will be filled faster than the last few.

Running the provided code

Testing

Start by running a simple end-to-end test.

cd mnist
python3 ml_test.py

This will pick a random digit from the MNIST test set, encrypt it, and evaluate the provided model on it using functional decryption.

Generating the confusion matrix

cd mnist
python3 confusion_matrix.py

This operates on cleartext data.

Benchmarking the scheme on your hardware

cd mnist
python3 -O benchmark.py

Project structure

• core: The code for the Quadratic FE scheme, optimized for 1 hidden layer Quadratic Networks.
  • discretelogarithm: A discrete logarithm solver that uses the Baby Step Giant Step method and interacts with a database to store and retrieve procomputations.
  • make_keys: A simple key generation function.
  • models: Classes and methods for vectors, cryptographic keys and machine learning models.
  • scheme: The implementation of the Quadratic Functional Encryption scheme.
  • utils: Defines some utilities, e.g. for serializing objects and batch exponentiations.
• mnist: Code testing and benchmarking the implementation on the MNIST dataset.
  • objects: A directory of directories where objects are stored durably. Should initially contain an instantiation (a choice of Pairing Group with base elements) and a pretrained MNIST model.
  • benchmark: Runs encryption and decryption on the whole MNIST test set and reports the timings. This takes a while.
  • confusion_matrix: Generates the confusion matrix found in the paper.
  • initialization: A run once script to run before running anything else. Generates keys and fills the database with discrete logarithm precomputations, which takes some time.
  • ml_test: Picks a MNIST digit at random, encrypts it and functionally decrypts it.
  • setup: Used to make importing from core easier.

Issues

Try opening an issue on GitHub.

Old paper

An earlier version of this work can be found on eprint. It may be of interest to people who seek a better understanding of this code or of the cryptography involved.