SIRNN: Secure Inference for Recurrent Neural Networks

SIRNN is an application of the SCI MPC protocols for inference over RNNs. This README discusses the inference over FastGRNN models.

Setup

For setting up SIRNN, set-up:

1. SCI
2. EzPC
3. SeeDot

Control Flow

SIRNN performs inference on quantized machine learning models. This quantization is carried out using the SeeDot framework. Seedot generates an output that is in the EzPC language. The EzPC compiler is then used to translate the code into SIRNN.

SIRNN code is essentially a sequence of function calls to the SCI library.

SIRNN Summary: SeeDot->EzPC->SCI.

Running Instructions

This section provided a step-by-step runthrough of SIRNN using the FastGRNN model and the Google-30 dataset.

Step 1: Training FastGRNN and Quantizing using SeeDot

1. Clone EdgeML and checkout sirnn branch:

git clone https://github.com/microsoft/EdgeML.git
git checkout sirnn

2. Follow the instructions in SeeDot's README to install SeeDot's dependencies.

3. Download Google-30 dataset.

cd EdgeML/examples/tf/FastCells/
mkdir Google-30/ && cd Google-30/
wget http://download.tensorflow.org/data/speech_commands_v0.01.tar.gz

tar -xvf  speech_commands_v0.01.tar.gz
cd ../

4. Process the dataset:

python process_g30.py -d Google-30/ -s Google-30/

5. Train FastGRNN model:

python fastcell_example.py -id 32 -hd 100 -rW 16 -rU 35 -sU 0.2 -sW 0.2 -dir Google-30/

6. Quantize using SeeDot and generate input:

cd ../../../tools/SeeDot/
mkdir -p model/fastgrnn/Google-30/
mkdir -p datasets/fastgrnn/Google-30/

cp -r ../../examples/tf/FastCells/Google-30/FastGRNNResults/<timestamp>/* model/fastgrnn/Google-30/

cp -r ../../examples/tf/FastCells/Google-30/train.npy datasets/fastgrnn/Google-30/
cp -r ../../examples/tf/FastCells/Google-30/test.npy datasets/fastgrnn/Google-30/

python fixSeeDotInput.py --seedot_file seedot/compiler/input/fastgrnn-g30.sd --model_dir model/fastgrnn/Google-30/ --dataset_dir datasets/fastgrnn/Google-30/ -n 1 --normalise_data

python SeeDot-dev.py -a fastgrnn -e fixed -d Google-30 -m red_disagree -t EzPC -n 1

export SEEDOT_DIR=$(pwd) # For accessing the files in the following steps

Warning: The above SeeDot run may fail in some devices due to a large number of threads being created. For resolution, reduce the number of rows in datasets/fastgrnn/Google-30/train.npy.

Step 2: Generating SIRNN code using EzPC

1. Clone EzPC repo:

git clone https://github.com/mpc-msri/EzPC.git

2. Install EzPC's dependencies and initialise OPAM environment.

3. Set directory environment variable for EzPC:

cd EzPC/EzPC/EzPC/

# Build EzPC compiler, if not already done.
make 

# The directory from where EzPC can be run.
export EZPC_DIR=$(pwd) 

# Folder for storing EzPC's output when run on SeeDot generated EzPC code. 

# This can be any convenient location
mkdir -p sirnn/  

4. Compile SeeDot generated EzPC code using the following steps:

cd ../../SIRNN/
mkdir FastGRNN/

python preProcessSIRNN.py --seedot_dir ${SEEDOT_DIR} --results_dir FastGRNN/ --predict_dir ${EZPC_DIR}/sirnn --dataset Google-30 --sci_build_location FastGRNN/build
python secureCodegen.py --predict_dir ${EZPC_DIR}/sirnn --ezpc_dir ${EZPC_DIR} --results_dir FastGRNN/Google-30/ 

mkdir -p FastGRNN/build/

Explanation: preProcessSIRNN.py script creates a folder with name 'dataset' in 'results_dir' folder. Then copies the SeeDot generated files to the 'dataset' folder. It copies the predict.ezpc file to a location indicated by 'predict_dir'. sci_build_location is the path that was specified for installing the SCI library. (If the location is not included within the $PATH environment variable.) securecppcodegne.py script runs the EzPC compiler on predict.ezpc, specified by 'predict_dir', and stored the output in 'sirnn_fixed.cpp' at 'results_dir'.

Note: The keywords in '' indicate the argument placeholders to the python scripts.

Step 3: Inference using SCI

1. Install SCI using the instructions in README.

   While installing SCI, the library should be installed to a location on the $PATH environment variable. It is preferable to use -DCMAKE_INSTALL_PREFIX flag and install the libraries to the location indicated by 'sci_build_location' above.

2. Compile SIRNN code

cd FastGRNN/build/
cmake .. # If compiling for the first time
make

3. Running SIRNN:

# For SERVER
./bin/Google-30 r=1 addr=0.0.0.0 port=8000 nt=4 inputDir=../Google-30/input/

# For CLIENT
./bin/Google-30 r=2 addr=0.0.0.0 port=8000 nt=4 inputDir=../Google-30/input/

Copyright (c) Microsoft Corporation. All rights reserved. Licensed under the MIT license.