VHAGaR:
In this repository, we provide an implementation for the paper "Verification of Neural Networks' Global Robustness" VHAGaR's paper. The repository owner is anan.kabaha@campus.technion.ac.il.

Prerequisites:

  Julia Version 1.11.3
  Gurobi 10.0 or 11.0 or 12.0 
  Python 3.8.10 
  Torch 2.4.1

Clone VHAGaR:

  git clone https://github.com/ananmkabaha/VHAGaR.git
  cd to VHAGaR
  run julia run.jl

VHAGaR parameters:

--dataset: the dataset, can be either mnist, fmnist, cifar10

--model_name: the model name, can be either FC0, FC1, CNN0, CNN1, or CNN2

--model_path: the path of the model

--perturbation: the peturbation can be either occ, patch, brightness, linf, contrast, translation, rotation, or max 

--perturbation_size: the size of the pertubation in this format: occ: i,j,width , patch: eps,i,j,width, brightness: eps, linf: eps, contrast: eps, translation: tx,ty, rotation: angle

--ctag: the source class

--ct: the target classes

--timeout: the MIP timeout

--output_dir: the output directory

Examples:

    julia run.jl --dataset mnist --model_name FC0 -m ./models/mnist_FC0/model.p --perturbation patch --perturbation_size 1,14,14,5 --output_dir ./results/ --ctag 1 --ct 2,3,4,5,6,7,8,9,10 --timeout 10800

    julia run.jl --dataset mnist --model_name CNN1 -m ./models/mnist_CNN1/model.p --perturbation occ --perturbation_size 1,1,5 --output_dir ./results/ --ctag 1 --ct 2,3,4,5,6,7,8,9,10 --timeout 10800