kNAVe (kNN Abstract Verifier) is an abstract interpretation-based tool for proving robustness and stability properties of kNN classifiers.
Given a training set D and a test set T, together with a perturbation P, kNAVe symbolically computes an over-approximation of P(x), the region of (possibly infinite) vectors that corresponds to feature variations of x in T, and runs a sound abstract version of the kNN on it, returning a superset of the labels associated with vectors in P(x). Clearly, when kNAVe only returns one label, kNN is provably robust and stable for such a perturbation.
- Python3
To install kNAVe you need to clone or download this repository and run the commands:
cd src
pip install ./
This will install the following dependencies:
- joblib
- nptyping
- numpy
- pandas
- pick
- python-dateutil
- pytz
- scikit-learn
- scipy
- six
- threadpoolctl
- tqdm
To run kNAVe:
cd src
python3 nave.py <config_file.ini>
where config_file.ini is located in the config folder, or alternatively:
cd src
python3 nave.py <config_file.ini> log
to obtain also a log file in the logs folder. For more information on usage and configuration file, run help.py without arguments.
Note: You can use ... to match multiple files. For example:
python3 nave.py ...
python3 nave.py str...
match, respectively, all files and files starting with str in the config folder.
Results are saved in 3 files:
- details.csv: contains classifications for all processed feature vectors
- robustness.csv: contains robustness results
- stability.csv: contains stability results
Their location depends on the specifications given in the configuration file.