mkskeller/idash-submission

This repository contains the scripts for our submission to track 4 of iDASH 2019. See our note for details.

Building the docker container

Dockerfile expects MP-SPDZ in the src directory as well as replicated-ring-party.x being compiled in said directory. You can either clone MP-SPDZ into src or fetch the compiled version from GitHub:

./get-mp-spdz.sh

Then build the container using

docker build .

Local executution

Training

Run

./train.sh <negative> <positive>

where <negative> and <positive> are text files containing the data, either split by comma or by tab. Furthermore, both files are expected to have a title row and column.

The script will prepare an optimized program for the virtual machines depending on the size of the data if such a program isn't present yet. Then it runs three virtual machines (replicated-ring-party.x) that execute the actual computation between them.

For simplicity, the data is all input via the first party (through src/Player-Data/Input-P0-0), after which all computation is entirely secret.

Prediction

Run

./predict.sh <data>

where <data> contains the input data in the same format as above. This will only work after running the training as above.

This will also run a secure computation and output the result as

predictions: <result>

where <result> is a string of 0s and 1s.

Remote executution

Docker setup

Run

docker run -it -p <ssh port>:2222 <vm port>:<vm port> mp-idash

on every host. This should ensure that all necessary ports are forwarded correctly and that that the SSH server is started (via .bashrc). Add -d to the options to run without active TTY (for example via SSH). Note that and have to the same and available on all hosts.

Training

On , run

./train-remote.sh <host0> <host1> <host2> <vm port> <ssh port> <negative> <positive>

where <negative> and <positive> are text files containing the data, either split by comma or by tab. Furthermore, both files are expected to have a title row and column.

The script requires that all hosts run an SSH server on and that is open. If running with docker the ports have to the same as above.

The script will prepare an optimized program for the virtual machines depending on the size of the data if such a program isn't present yet. This optimized program is distributed by rsync together with the SSL certificates. Then it runs three virtual machines (replicated-ring-party.x) that execute the actual computation between them.

For simplicity, the data is all input via the first party (through src/Player-Data/Input-P0-0), after which all computation is entirely secret.

Prediction

Run

./predict-remote.sh <host0> <host1> <host2> <vm port> <ssh port> <data>

where <data> contains the input data in the same format as above. This will only work after running the training with train-remote.sh, and it also requires SSH and open ports as above.

This will run a secure computation as for training and output the result as

predictions: <result>

where <result> is a string of 0s and 1s.

Computing the accuracy

We have used regression.mpc as provided with MP-SPDZ 0.1.2 for our figures. You can compile it as follows in the root directory of MP-SPDZ:

./compile.py -DR 64 regression <n_epochs> <n_threads> [opts]

where <n_epochs> stands of the number of epochs and <n_threads> stands for the number of threads being used. Further options are:

• bc for the BC-TCGA dataset (default is GSE2034)
• nearest for nearest rounding (default is probabilistic truncation)
• tol=<x> to stop when the loss is below x (default is fixed number of epochs)

regression.mpc processes inputs differently than the above programs. It expects first the labels for all examples, then every feature for all examples, all through player 0. In other words, Player-Data/Input-P0-0 should contain a whitespace-separated table where first row is for labels and the further rows are the features.

Note that the program computes a non-standard aggreated accuracy measure, but you can compute any accuracy measure using the two a/b outputs on lines starting with test_acc: (negative first).