Paper: SORA: Scalable Black-box Reachability Analyser on Neural Networks [link]
All experiments are carried out on a workstation equipped with 96GB RAM, a 20-core Intel i9-10900X CPU, and an Nvidia 2080Ti GPU.
Requirement:
PyTorch==1.11.0
torchvision==0.12.0
numpy
We used the following pretrained models provided in ERAN:
mnist_relu_6_100
mnist_relu_6_200
mnist_relu_9_200
ffnnRELU__Point_6_500
convSmallRELU__Point
convMedGRELU__Point
mnist_conv_maxpool
convSuperRELU__DiffAI
The original models are with the .onnx format, so we convert them to fit the PyToch.
The converted version can be download from google drive. [link]
Two shell scripts are provided to run the experiments in our paper. Please check and update the path variables in these scripts before using them.
mkdir model
# Download those pretrained models and put them in `model/`
sh run_go.sh
sh run_sora_pgd.sh