• OS: Ubuntu 20.04

• Pytorch: 1.12.1+cu113

• Torchvision: 0.13.1+cu113

We provide the training code to obtain a well trained DNN model weight parameters. The supported network architectures and tasks are listed as following:

Networks:

• ResNet-34
• ResNet-50
• VGG-16

Datasets:

• CIFAR-10
• GTSRB
• ImageNet [Only used for inference]

Step 1: go into train_on_torch folder.

cd infras

Step 2: prepare training data. To build an unified training and test data read and write format. We use ImageFolder API provided by torchvision.datasets to save all of our datasets as .png images. With the followed structure:

--cifar10
    --train
        --0
        --1
        --...
        --9
    --test
        --0
        --1
        --...
        --9

The functionality is implemented in prepare_datasets.py, and just run

python prepare_dataset.py --dataset cifar10

After this step, you should have a re-arranged folder to ImageFolder to read image data, and convert them as dataloader.

Step 2: run train.py to obtain a well trained DNN model weight parameters.

python train.py --network vgg16 --dataset gtsrb

The trained DNN model parameters are stored in ../trained_models folder. The trained pytorch weight parameters are named as best.task.3.ckpt and storaged according to its network and dataset in ../trained_models.

Step 3: we need export the trained model parameters as .txt format. This is impelemented in file export_model.py.

python export_model.py --network vgg16 --dataset gtsrb

After that, you will export the model parameters as float and storaged in ../trained_models/gtsrb_vgg16/best.task.txt.

• gcc : gcc (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0

• make: GNU Make 4.2.1

• OpenBLAS: 0.3.20

1. step into the poc folder

cd poc

2. run make to compile and install openblas-0.3.20

make install

In this step, the exported model parameters will be used to inference by our ML inference Infrastructure which is implemented in folder poc/infras. For example, if you want to run gtsrb task on vgg16 network, you can perform:

cd poc/infras
make gtsrb_vgg16

Then, the exported model parameters will loaded in and evaluate on the same testset.

This infrastructure provides a tesebed to evaluate our fualt injections on OpenBLAS library.

All of our attacks, including LLVM-VIS algorithms and rowhammer exploitations are evaluated on this infrastructure.

• cmake: 3.16.3

• clang: ubuntu clang version 12.0.0-3ubuntu1~20.04.5

cd ~
wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh 12

you need to change your current soft link of clang, clang++, opt to your installed above:

clang --version
which clang
ls -l /usr/bin | grep clang
ln -s /usr/lib/llvm-12/bin/clang /usr/bin/clang

repeat the same process to update the soft link of clang++, opt.

The core function of our LLVM-VIS algorithm is implemented as a llvm-pass (attack\source\ground_truth\llvm-pass\flipBranches and attack\source\ground_truth\llvm-pass\flipLauncher). So, at first, we need to complie this llvm pass as a .so library.

Before that, please check whether your cmake version is ```3.16.3``.

There are two absolute path in file flipLauncher\flipLauncher.cpp should be replaced as your own absolute path.

• In line of 20, #define BRANCH_INFO_FILE "/home/xxx/SGXBLAS/attack/source/ground_truth/br_info.tmp"

• In line of 79, std::string command_prefix = "opt -enable-new-pm=0 -load ${HOME}/xxx/SGXBLAS/attack/source/ground_truth/llvm-pass/build/flipBranches/libflipBranches.so -flipbranches -o ";

After check them, we start compile our llvm-pass:

cd attack\source\ground_truth\llvm-pass
mkdir build && cd build 
cmake ..
make 

Put all together, we implement our attacks in attack\source\ground_truth\Makefile, the only you need to do is specific the attack instance, e.g., gtsrb_vgg16 in the line of attack\source\ground_truth\Makefile file.

Before that, you need check the path of our compiled llvm, in particular, you need check the following files:

ground_truth/openblas_makefiles/attack/dirvier/level3/Makefile

ground_truth/openblas_makefiles/attack/interface/Makefile

please replace the path ${HOME}/xxx/SGXBLAS in above files as your own path.

Before attacking, you need to specific your attack instance in attack\source\ground_truth\Makefile in line 34, 35, 36:

EXP_CONF = gtsrb_vgg16
EXP_DATASET = gtsrb
EXP_NETWORK = vgg16

Finally, we have it!

cd attack\source\ground_truth
make ground_truth