This tool implements the evaluation framework described in the paper
On (The Lack Of) Location Privacy in Crowdsourcing Applications by Spyros Boukoros (TU Darmstadt), Mathias Humbert (SDSC, ETHZ/EPFL), Stefan Katzenbeisser (TU Darmstadt/U. Passau), and Carmela Troncoso (EPFL). USENIX Security Symposium 2019
Link to the paper: https://arxiv.org/abs/1901.04923
The main files that need tweaking are the
- __init__.py
- pipelineJson.json
You have to change the path in the section "Required paths for applying defenses" __init__.py#L24
mypath: the path to the folder of your project
original_csv_file: the path to your dataset. This file needs to be called original.csv
priors_csv: the path to the dataset that will be used for calculating the prior probabilities of locations.
This file is relevant only for defenses that are augmented with optimal remapping.
This file needs to be called data_for_priors.csv.
This file includes parameters abou the attacks, and the datasets used.
If you wand to run the tool on all of your dataset, then place needs to be call "World.". Otherwise, you can change it to whatever you want however, then you need to specify a valid polygon for the extraction.
Detailed:
Place: World if no region extraction otherwise free to choose
dataset_used: the name of you dataset. In this tool, we provide two datasets and their utility functions. Those are the "Safecast" and
"Radiocells" datasets. The variable name is case sensitive. If you specify your own dataset name, you will not be able to use the utility
functions of these datasets. However, you can tweak the tool to include your own funtion.
minpoints: the DBSCAN's parameter minimum points
maxdistance: DBSCAN's parameter maximum distance among points
maxpois: The top N clusters that should be kept
starttime: Keep only measurement that start from this time
endtime: and end this time. Basically, the last two parameters filter the dataset.
advanced_clustering: Whether to use standard clustering parameters (value=0) or adjustable(value=1).
min_allowed_points: When using adjustable clustering, the minimum limit points are allowed to reach.
max_allowed_distance: When using adjustable clustering, the maximum distance allowed.
original_min_points: When using adjustable clustering, the value can replace the initial minpoints.
Polygon: This needs to be a valid (closed) polygon and the edges need to be in the form of (latitude, longitude).
See here for various polygons --> https://download.geofabrik.de/. However, the above link provides the polygons in the opposite order (longitude, latitude). If you use such polygons make sure you reverse the edges.
The main functions that automates everything is inside the
automated_pipeline.py file and is called runpipeline.
python automated_pipeline.py
When you call runpipeline, the tool
- extracts the region,
- computes the priors if any,
- apply all defenses to the location data.
- The tool saves the defense files in a folder called defenses.
You can select which defenses you want in the file apply_defenses.py.
The tool calculates all users original privacy in spatial terms and POIs.
In order to do so, rellies on the attackscript inside the attackmodule.py. If you wish to change the attackm, modify this file.
In addition, the tool relies on OSM to collect POIs. Currently it uses a public server that if under load starts blocking requests. The tool has defenses for this and tries again after x seconds.
However, you could also use your own OSM server by modifying the line api = overpy.Overpass(url='your url') inside the privacy_gain.py file.
Then, the tool applies the attacks also on the defense files. All resulting files are pickled and saved in the privacyloss folder.
The tool, if Safecast or Radiocells are specified (and correct files are provided), calculates the utility loss.
As the tool relies on a specific order of the files it finds, (the columns in the csv should be in specific order), we provide templates for the Radiocells and the Safecast datasets. The graphs are saved in the plots folder.
The tool calculates the privacy gains and plots it. Again, this is a demonstration on the datasets of the paper.
The privacy gain folder with each user's privacy gain is saved in the privacyloss folder. The graphs however are stored in the plots folder.