Code for the eperimental work as part of the paper: Visibility Graph-Based Wireless Anomaly Detection for Digital Twin Edge Networks, IEEE Open Journal of the Communications Society, 2024.
It has been recently demonstrated that Graph Neural Network (GNN)s need only a fraction of the trainable weights to achieve similar performance as more mature Deep Learning (DL) methods. Additionally, GNNs are also capable of capturing the dynamic relationships among data points in a Time Series (TS) and excel in mapping intricate interactions and connections within the attributes of each individual data point. Inspired by this, in this paper we propose a computationally efficient DITEN monitoring approach for wireless link-layer anomaly detection based on a time-series to graph transformation and GNNs.
The Rutgers.ipynb notebook is designed to convert time series data from the Rutgers dataset into graphs. These graphs are categorized into 5 groups, each representing a specific anomaly type. The notebook includes a GINE model for classifying the graphs based on the anomaly types and a GAT model for identifying the nodes where the anomalies occurred within the time series. The datasets were created with Rutgers_dataset_creation.ipynb.
Rutgers_explainability.ipynb is a notebook that builds on Rutgers.ipynb. It calculates importance scores for each node using the gradient method. The importance score can be calculated when the model is trained by sending the test loader through the model and inspecting the gradients. Nodes with higher scores are considered more important. To assess the accuracy of the importance scores, the notebook includes a modified test step where the most important nodes are masked, and the prediction changes are observed. You can choose between the importance calculation method provided by GraphXAI and the one implemented in this notebook.
All notebooks were run in jupiter hub, except Rutgers_dataset_creation.ipynb that was run in google colab.
Each of the three main notebooks contains cells at the beginning that install the required versions of the packages.
In each notebook, the cells, with the exception of pip install, Config, and the last main cell, are functions that must be executed before the last cell is executed, which runs the entire program.
To change the parameters, you can edit the Config Cell located before the last cell in each notebook.
@ARTICLE{10508580,
author={Bertalanič, Blaž and Hribar, Jernej and Fortuna, Carolina},
journal={IEEE Open Journal of the Communications Society},
title={Visibility Graph-Based Wireless Anomaly Detection for Digital Twin Edge Networks},
year={2024},
volume={5},
number={},
pages={3050-3065},
doi={10.1109/OJCOMS.2024.3393853}}
This work was supported by the European Commission NANCY project (No. 101096456)