tudelft-cda-lab/FATE

Fuzzing for Adversarial examples in Tree Ensembles

Fuzzing for adversarial examples

1. Make sure you have installed libFuzzer (clang-11) and optionally the LT-attack (executable should be in the root of this repo as lt_attack), the Gurobi optimizer, AFL++ and honggfuzz
2. Set the appropriate paths in python/constants.py
3. Download the data.
   • breast-cancer, diabetes, ijcnn1, covertype from LibSVM: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/.
   • higgs 90.000 instances from OPENML: https://www.openml.org/search?type=data
   • (F)MNIST: original datasets
   • Make sure the data is put in the right location (inspect python/load_data.py)
   • Other datasets can be added easily. Just specify them in python/datasets.py and python/load_data.py
4. Adjust the settings in python/constants.py according to your wishes
5. There are 2 command line tools: python/run_fate.py and python/automate.py. Execute with -h flag for detailed descriptions and options.
6. Run python3 python/run_fate.py [dataset_name] [model_type] --reload once for each dataset
7. Using --reload is only necessary for the first run for each dataset. It will generate/train the model to be fuzzed. Afterwards, do not provide --reload for speedup (training the model is not necessary anymore) and consistency.
8. Optionally, the baseline methods can be run via python/run_milp.py and python/run_zhang.py
9. Most settings in python/constants.py (everything that does not have to do with the training of the model) can be adjusted without having to specify --reload again.