Digital Twin of the Radio Environment: A Novel Approach for Anomaly Detection in Wireless Networks

This code is used for the paper:

A. Krause, M. D. Khursheed, P. Schulz, F. Burmeister and G. Fettweis, "Digital Twin of the Radio Environment: A Novel Approach for Anomaly Detection in Wireless Networks," in Proceedings of 2023 IEEE Globecom Workshops: 3rd Workshop on Sustainable and Resilient Industrial Networks (GC 2023 Workshop - SRINetworks 2023), Kuala Lumpur, Malaysia, Dec 2023.

A preprint of the paper can be found on arXiv. The paper on IEEE Xplore can be found here.

Dataset:

Workflow

The workflow is split into two parts: dataset generation and anomaly detection on the generated data.

Data Formats

The following data formats are used in the process of results generation. fspl indicates that the free-space path loss model was to used to generate the path loss maps.

• Path loss map: Contains a collection of path loss maps. File name is fspl_PLdataset<nr>.pkl. Each path loss map indicates the path loss between a specific transmitter location and each pixel on the map in dB.
• Radio map: Contains a collection of radio maps. File name is fspl_RMdataset<nr>.pkl. Each radio maps combines one or several transmitters (regular or jammer) with a specific power and path loss and adds everything up, so that the result is the total RSS at each pixel on the map in dBm.
• Measurements: Contains a collection of measurements. File name is fspl_measurements<nr>.pkl. Each measurement entity is a collection of values, whereby each value is the difference between the measured (original) RSS and the RSS expected from the digital twin at the same location.
• Results: Contains the results of anomaly detecion for a given measurement data set. File name fspl_results<nr>.pkl. Each file contains a dictionary with the following entries: y_test, y_hat and jammer. jammer refers to an array, in which the transmitters of type jammer are saved or an empty array, respectively in case there is no jammer. y_hat is a score to allow a soft decision / ROC curve creation.

Dataset Generation

Note: The dataset is uploaded at https://zenodo.org/records/10512316. Instead of generating the dataset, you can also download it and save it in the datasets folder. Then, you can skip the dataset generation and directly execute the anomaly detection.

The dataset generation consists of three steps. For each script, there is a corresponding yaml file in the conf folder, which contains the parameters for each step. For each dataset file, a .txt file is generated, which contains the parameters used for the generation of the dataset. The dataset files are saved in the datasets folder.

1. Create pathloss map dataset using src\dataset_generation\pathloss_map_generation.py. In this step, for random transmitter positions path loss maps including shadowing are generated.
2. Create a radio map dataset using src\dataset_generation\radio_map_generation.py. First, the number of regular transmitters $N_{reg}$ is specified. Then, one pathlossmap is randomly chosen for each transmitter (i.e., random locations of the transmitters). Each pathloss map is converted into a radiomap by taking the transmit power into account($P_{rx} = p_{tx} - L). The received power are added up to obtain the total received power at each pixel on the map. For some maps (according to the specified jammer probability), a jammer is added to the map. The jammer is also based on one of the pre-generated pathloss maps.
3. Create a measurement dataset using src\dataset_generation\measurement_generation.py. In this step, for each of the radio map from step 2, a digital twin radio map is generated (see Section IV.A in the paper for more information). Subsequently, from the sensing unit positions the RSS values are extracted from both the physical twin and the digital twin radio map and the difference is calculated.

Typically, there is one pathloss map dataset with a given shadowing variance. From this, one radio map dataset is generated for a specified number of transmitters. From the radio map different measurement datasets for different grid sizes / number of sensing units can be generated.

Anomaly Detection

The script src\anomaly_detection\fspl_anomaly_detection.py executes the anomaly detection. The script takes the measurement datasets as input. For each measurement dataset, the anomaly detection is executed and the results are saved in a results file. The results file contains for each method and each sample a score/probability, whether the sample is an anomaly or not (as long as in fspl_anomaly_detection.yaml the parameter probability is set to True).

In the file fspl_anomaly_detection.yaml in the conf folder, the parameters for the anomaly detection can be specified. The most important parameters are:

• vary_parameter: The parameter over which is iterated when executing fspl_anomaly_detection.py. Can be either noise_std or grid_size.
• method: The anomaly detection algorithm.

You can find more information in the yaml file.

After the results are generated, they can be visualized using the Jupyter notebook notebooks\roc_curves_fspl.ipynb.

Reproducing the Results

Due to the given size of the dataset it is not uploaded. You can either contact me to get the dataset or generate it yourself. The following steps are necessary to reproduce the results:

1. The pathloss maps need to be reproduced. Herefore, execute step 1 from the section Dataset Generation. The yaml file for this step is conf\pathloss_map_generation.yaml. The following parameter combinations need to be run:
   • dataset_nr: 0, 1, 2, 3, 4, 5
   • fspl.noise_std: 0, 2, 4, 6, 8, 10
2. Execute the second step from the section Dataset Generation. The yaml file for this step is conf\radio_map_generation.yaml. The parameter list for execution is the following:
   • dataset_nr: 0, 1, 2, 3, 4, 5
3. Execute the third step from the section Dataset Generation. The yaml file for this step is conf\measurement_generation.yaml. The parameter list for execution is the following:
   • grid_size: 10, 10, 10, 10, 10, 10, 5, 15, 20
   • rm_dataset_nr: 0, 1, 2, 3, 4, 5, 1, 1, 1
   • meas_dataset_nr: 0, 1, 2, 3, 4, 5, 10, 11, 12

In the last step, the datasets are generated for a grid size of 10m with the shadowing noise standard deviations [0, 2, 4, 6, 8, 10]. Additionally, the datasets are generated for a shadowing noise standard deviation of 2dB with grid sizes of [5, 15, 20]m. (The dataset for 10 m is already generated.)

Methods for Anomaly Detection

Conventions

• Binary classification:

  • 0: Normal
  • 1: Anomality

• The indexing is [x, y], even though numpy uses matrix indexing [i, j]. This has to be considered when plotting.

Unsupervised Methods

The best performing unsupervised methods are:

• One-class SVM
• Local outlier factor
• Adapted energy detector (in the code, it is referred to as unsupervised_threshold)

Yet, there are more unsupervised and even supervised methods for comparison implemented. Please note, that they may not be fully mature.

Other stuff

Positioning inaccuracy

See paper > IV.A Building the Digital Twin for more information on how the poisitioning inaccuracy is modeled.

Citation

If you use the code or the dataset, please cite the following paper:

@INPROCEEDINGS{krause2023digital,
author={Krause, Anton and Khursheed, Mohd Danish and Schulz, Philipp and Burmeister, Friedrich and Fettweis, Gerhard P.},
title={Digital Twin of the Radio Environment: A Novel Approach for Anomaly
Detection in Wireless Networks},
booktitle={2023 IEEE Globecom Workshops (GC Wkshps): 3rd Workshop on Sustainable and
Resilient Industrial Networks (GC 2023 Workshop - SRINetworks)},
address={Kuala Lumpur, Malaysia},
month={12},
year={2023},
}