COINN: Crypto/ML Codesign for Oblivious Inference via Neural Networks

This repo contains the executables for benchmarking the runtime for oblivious inference through COINN. The executable is run inside a docker.

Start docker

docker pull siamumar/ubuntu18.04-coinn:first
docker run -it --cap-add=NET_ADMIN ubuntu18.04-coinn /bin/bash
cd home/_COINN_

Download necessary files

inside /home/_COINN_, clone the COINN repository:

git clone https://github.com/ACESLabUCSD/COINN.git
cp -rf COINN/bash_files .

A folder named bash_files will appear, which contains the commands needed to run for each neural network benchmark in the paper.

Simulate LAN and WAN bandwidth

Run one of the follwing commands to simulate LAN or WAN bandwidth on localhost

bash simulate_lan.sh # LAN
bash simulate_wan.sh # WAN

Split the terminal to run server and client

Use tmux to split the teminal in two as follows:

• run tmux
• press ctrl + b, then shift + quotation mark ("). the terminal will be split in two.
• press ctrl + b, then up or down arrows to navigate between splitted terminals.

Run oblivious inference benchmarks

In the following commands, replace <NETWORK> with one from the list: cifar10-BC1, cifar10-resnet110, cifar100-resnet32, imagenet-resnet50.

In the server terminal, run

cd bash_files/<NETWORK>/server
bash commands_setup.sh # the first time inference is run
bash commands.sh # from the second time on

In the client terminal, run

cd bash_files/NETWORK/client
bash commands_setup.sh > ../../../timing_reports/<NETWORK>.txt # the first time inference is run
bash commands.sh > ../../../timing_reports/<NETWORK>.txt # from the second time on

The layer-wise runtimes are written to timing_reports/<NETWORK>.txt.

Summarize runtimes

After running all networks, run the following to summarize the runtimes.

python3 summarize_results timing_reports

The summary is written to timing_reports/summary.csv