NeuroShield

NeuroShield is an end-to-end DNN compilation toolchain that applies dedicated obfuscation primitives for compiled DNN binaries. NeuroShield is built upon Apache TVM

How to run

This code base is tested on Ubuntu 22.04, llvm-14, Python 3.9. Install dependencies by:

$ sudo apt install llvm-14 llvm-14-dev
$ cd <path2thisrepo>
$ pip install -U setuptools # we recomend use virtualenv
$ pip install -r requirements.txt
$ pip install numpy==1.19.5 # ignore the error

Build tvm

init:

$ cd <path2thisrepo>
$ git submodule update --init

apply patch:

$ cd <path2thisrepo>/tvm
$ git submodule update --init
$ git apply ../dnnobf.patch

change the setting in <path2thisrepo>/tvm/cmake/config.cmake build:

set(USE_MICRO ON)
set(USE_LLVM /usr/lib/llvm-14/bin/llvm-config)

build tvm:

$ cd <path2thisrepo>/tvm
$ mkdir build && cd build
$ cp ../cmake/config.cmake .
$ cmake .. -G Ninja
$ ninja

build tvm python:

$ export TVM_HOME=<path2thisrepo>/tvm
$ export PYTHONPATH=$TVM_HOME/python:${PYTHONPATH}
$ cd <path2thisrepo>/tvm/python
$ python setup.py install

Build DNN binaries

build llvm helper pass:

$ cd <path2thisrepo>/bingen/llvm_pass
$ mkdir build && cd build
$ cmake ..
$ make

generate model onnx file (e.g., mnist):

$ cd <path2thisrepo>/models
$ python generate_onnx.py mnist

profile model neuron ranges:

$ cd <path2thisrepo>/profile
$ python neuron_profile.py mnist train 

Note: it takes about 3 days to profile all six models on an Intel Xeon Gold 6258R machine; this step can be skipped, in that case NeuroShield will generate a large enough neuron range

modify <path2thisrepo>/bingen/Makefile settings:

model_name := mnist # resnet, mobilenet, fasttext, esm, albert
dnnobf_flex_fuse := 1 # set to 0 to disable ff
dnnobf_fakeop_insert := 1 # set to 0 to disable fi
dnnobf_compute_reorder := 1 # set to 0 to disable cr

generate DNN binaries:

$ cd <path2thisrepo>/bingen
$ make

run generated binaries (e.g., mnist)

$ cd <path2thisrepo>/bingen/build_output/ff{dnnobf_flex_fuse}_fi{dnnobf_fakeop_insert}_cr{dnnobf_compute_reorder}
$ ./dnn_bin mnist_input.bin