Modular Neural Networks for Anomaly Detection in Cyber-Physical Systems

This is the repository for our publication Using Modular Neural Networks for Anomaly Detection in Cyber-Physical Systems.

Overview

We use Modular Neural Networks to model the inner dependencies of Cyber-Physical System (CPS) subsystems. Thereby, we can achieve a more robust detection of anomalies in CPS and a better allocation of their root-causes. For further information, we recommend you to read our [publication(#citation)].

Table of Contents

• Requirements
• Replication
• Citation
• License

Requirements

We recommend Anaconda to install all requirements for our repository. The requirements are saved in the venv.yml file.

For a quick installation run: conda env create -f venv.yml

Replication

As empirical validation dataset, we used the robot-anomaly dataset of Grabaek et al. 2023. You can access and download the dataset here. Once you downloaded the dataset, save it in the ./data directory.

For replicating the results from our paper, run the main.py script from the ./code directory. The script will run the reproducible hyperparameter search, as well as the subsequent replication studies, and evaluation studies.

By running the ./exp/exp_setup/evaluation.ipynb, you can calculate the metrics from the paper.

You can run your own studies by uncommenting the suitable codeblock in the main.py script.

Citation

When using this work, please cite:

@inproceedings{Ehrhardt2024,
    title={Using Modular Neural Networks for Anomaly Detection in Cyber-Physical Systems},
    author={Ehrhardt, Jonas and Overlöper, Phillip and Vranjes, Daniel and Steude, Henrik and Diedrich, Alexander and Niggemann, Oliver},
    year={2024},
}

LICENSE

Licensed under MIT license