Outside the Box

This repository contains the implementation and data used in the paper Outside the Box: Abstraction-Based Monitoring of Neural Networks, published at ECAI 2020. To cite the work, you can use:

@inproceedings{outsidethebox19, author = {Thomas A. Henzinger and Anna Lukina and Christian Schilling}, editor = {Giuseppe De Giacomo and Alejandro Catal{\'{a}} and Bistra Dilkina and Michela Milano and Sen{\'{e}}n Barro and Alberto Bugar{\'{\i}}n and J{\'{e}}r{\^{o}}me Lang}, title = {Outside the Box: Abstraction-Based Monitoring of Neural Networks}, booktitle = {{ECAI}}, series = {Frontiers in Artificial Intelligence and Applications}, volume = {325}, pages = {2433--2440}, publisher = {{IOS} Press}, year = {2020}, url = {https://doi.org/10.3233/FAIA200375}, doi = {10.3233/FAIA200375} }

Installation

We use Python 3.6 but other Python versions may work as well. The easiest is to use a conda environment. The package requirements that need to be installed are found in the file requirements.txt.

Recreation of the results

Below we describe how to obtain the results shown in the paper.

Data

Due to Gihub's limitation of the file size, one of the datasets needed to be compressed. To use this dataset, you need to manually go to the folder data/GTSRB/ and unzip the file train.zip.

Models

The repository contains the pretrained models used in the evaluation. The models can be recomputed using the scripts run/train_INSTANCE.py where INSTANCE is the name of the model/data combination.

Evaluation

The scripts to reproduce the figures and tables of the paper are found in the folder run/:

• Fig. 2: plot_toy_model.py
• Fig. 3: plot_boxes.py (Note that this will not produce the exact same figure because we obtained the figure for a network with different training parameters that we forgot to note down.)
• Fig. 4: plot_explanation_alpha_thresholding.py
• Table 2: plot_legend.py
• Figures 5-8 and Table 3: run_experiments.py (This script calls scripts for the individual experiments that can also be run in isolation.)

Intermediate results of the experiments are stored in .csv files in the run/ folder. The final plots are stored in the top folder.