MC²: A Platform for Secure Analytics and Machine Learning

Born out of research in the UC Berkeley RISE Lab, MC² is a platform for running secure analytics and machine learning on encrypted data. With MC², users can outsource their confidential data workloads to the cloud, while ensuring that the data is never exposed unencrypted to the cloud provider. MC² also enables secure collaboration, i.e., multiple data owners can use the platform to jointly analyze their collective data, without revealing their individual data to each other.

MC² provides the following (actively maintained) secure computation services:

• Opaque SQL: Encrypted data analytics on Spark SQL using hardware enclaves
• Secure XGBoost: Collaborative XGBoost training and inference on encrypted data using hardware enclaves
• Federated XGBoost: Collaborative XGBoost in the federated setting

The MC² project also includes exploratory research prototypes that develop new cryptographic techniques for secure computation. Please visit the individual project pages for more information:

• Cerebro: A general purpose Python DSL for learning with secure multiparty computation.
• Delphi: Secure inference for deep neural networks.

This repository contains the source code for the MC² client, which enables users to easily interface with MC² services deployed remotely in the cloud. Currently, the client supports remote deployments of Secure XGBoost and Opaque SQL only. To run an end-to-end MC² workflow:

1. Launch Opaque SQL or Secure XGBoost in the cloud (instructions to do so can be found in the respective repositories)
2. Use the MC² client to encrypt data locally, transfer it to the cloud VMs, run scripts specifying the desired computation, and retrieve and view encrypted results.

Alternatively, to use the individual services without the MC² client, please visit the respective project pages linked above.

Table of Contents

• Quickstart
• Documentation
• Contact

Quickstart

To quickly get a flavor of MC², you can work in a Docker image that comes with pre-built versions of MC² Client, Opaque SQL, and Secure XGBoost, and all dependencies. This quickstart is completely self-contained within a container.

1. You must have Docker installed. We recommend giving Docker at least 2 CPUs, 6 GB of memory, and 2 GB of swap space (instructions for Mac, Windows). Without sufficient resources, the quickstart may not work.

   Once that is done, pull the Docker image and launch a container.

   docker pull mc2project/mc2
   docker run -it -p 22:22 -p 50051-50055:50051-50055 -w /root mc2project/mc2

   Start an SSH server inside the container. (Note that this step is Docker-specific, and you only need to start the SSH server if running in a Docker container -- most cloud VMs automatically start an SSH server on boot.)

   service ssh start
   

2. Navigate to the mc2-client/demo directory. The configuration for this quickstart has been pre-populated in demo/mc2.yaml. More on the configuration can be found here. By default, the configuration has been set assuming you want to run Secure XGBoost. If you want to run Opaque SQL instead, comment out the Secure XGBoost section in the local part of the YAML configuration and comment in the Opaque SQL section.

   # Configuration for local data
   local:
       # If you want to run Secure XGBoost
       # Your data to compute on
       data:
           - data/securexgb_train.csv
           - data/securexgb_test.csv
   
       # Secure XGBoost script to run
       script: secure_xgboost_demo.py
       # ----------------------------------
   
   
       # # If you want to run Opaque SQL
       # # Your data to compute on
       # data:
       #     - data/opaquesql.csv
       # 
       # schemas:
       #     - data/opaquesql_schema.json
       # 
       # # Opaque SQL script to run
       # script: opaque_sql_demo.scala
       # # ------------------------------

   Included in the repo are a sample Secure XGBoost script and a sample Opaque SQL script.

3. Start the desired compute service within the container (Secure XGBoost or Opaque SQL). In a production environment, these compute services would be started in the cloud. Starting a compute service will start a listener that listens on port 50052.

   # Start the Secure XGBoost service. Replace `--xgb` with `--sql` for Opaque SQL instead.
   mc2 launch --xgb

   The Secure XGBoost service will take a few seconds to start, while the Opaque SQL service will take anywhere between 20-30 seconds to start. You can check whether the service is ready:

   mc2 check

4. Next, encrypt and transfer the encrypted data. Data to be encrypted/transferred is in mc2.yaml (this is pre-populated with the sample data). In this quickstart, the "transfer" is just a scp to another directory in the same container. In practice, the transfer is an upload to a remote machine in the cloud. The destination path for the data can also be specified in the configuration YAML under cloud/data_dir. In the demo directory, run the following command depending on which compute service you've started.

   cd mc2/demo
   
   # Specify the `--xgb` flag if running Secure XGBoost. Specify `--sql` for Opaque SQL instead.
   mc2 upload --xgb

5. Now, you're ready to run computation. Start computation through MC² according to the compute service.

   # Specify the `--xgb` flag if running Secure XGBoost. Specify `--sql` for Opaque SQL instead.
   mc2 run --xgb

6. Once computation has finished, download results. The source and destination of downloaded results can be specified in the configuration YAML under cloud/results and local/results, respectively. To also decrypt results, specify either --xgb or --sql to decrypt results outputted by Secure XGBoost or Opaque SQL, respectively.

   For this quickstart, the predictions outputted by Secure XGBoost are sent over the network and automatically decrypted client-side instead of saved to a file, so you will not need to decrypt results if running Secure XGBoost.

   # Download results
   mc2 download
   
   # If running Opaque SQL, download results and decrypt them
   # mc2 download --sql

Documentation

For more thorough documentation on installation and usage, please visit:

• MC² Client
• Opaque SQL
• Secure XGBoost

Contact

For questions and general discussion, please reach out to mc2-dev@googlegroups.com.