This repository contains the Python code for applying different k-anonymisation algorithms, i.e., Optimal Lattice Anonymization (OLA), Mondrian, Top-Down Greedy Anonymisation (TDG), k-NN Clustering-Based (CB) Anonymisation, on datasets and measuring their effects on Machine Learning (ML) classifiers as presented in k-Anonymity in Practice: How Generalisation and Suppression Affect Machine Learning Classifiers.

@article{slijepvcevic2021k,
  title={k-Anonymity in practice: How generalisation and suppression affect machine learning classifiers},
  author={Slijep{\v{c}}evi{\'c}, Djordje and Henzl, Maximilian and Klausner, Lukas Daniel and Dam, Tobias and Kieseberg, Peter and Zeppelzauer, Matthias},
  journal={Computers \& Security},
  volume={111},
  pages={102488},
  year={2021},
  publisher={Elsevier}
}

In order to install the necessary requirements either use pipenv install or pip3 install -r requirements.txt. Then activate the virtual environment, e.g. with pipenv shell.

The code is written in Python 3 and conducts following steps for each experiment:

• read specified dataset
• measure specified ML algorithm performance using original dataset
• anonymise dataset with specified algorithm and current value k for k-anonymity
• measure specified ML algorithm performance using anonymised dataset
• repeat previous steps for other configured values of k

The parameters, i.e., dataset, ML algorithm, k-anonymisation algorithm, and k are defined via arguments as follows:

usage: baseline_with_repetitions.py [-h] [--start-k START_K] [--stop-k STOP_K] [--step-k STEP_K] [--debug] [--verbose] [{cmc,mgm,adult,cahousing}] [{rf,knn,svm,xgb}] {mondrian,ola,tdg,cb} ...

Anonymize data utilising different algorithms and analyse the effects of the anonymization on the data

positional arguments:
  {cmc,mgm,adult,cahousing}
                        the dataset used for anonymization
  {rf,knn,svm,xgb}      machine learning classifier
  {mondrian,ola,tdg,cb}
    mondrian            mondrian anonyization algorithm
    ola                 ola anonyization algorithm
    tdg                 tdg anonyization algorithm
    cb                  cb anonyization algorithm

optional arguments:
  -h, --help            show this help message and exit
  --start-k START_K     initial value for k of k-anonymity
  --stop-k STOP_K       last value for k of k-anonymity
  --step-k STEP_K       step for increasing k of k-anonymity
  --debug, -d           enable debugging
  --verbose, -v

The k-anonymisation algorithms "k-NN Clustering-Based Anonymisation", "Mondrian" and "Top-Down Greedy Anonymisation" located in the folders clustering_based, basic_mondrian and top_down_greedy are based on the open-source implementation of Qiyuan Gong.

The original reporitories can be found on github.com:

• Clustering Based k-Anonymization
• Basic Mondrian
• Top Down Greedy Anonymization

Our changes include the migration of Python 2 to Python 3, the option to leave non-QID attributes and the target variable non-anonymised, the ability to handle float numbers in datasets, removal and cleanup of files and code that were irrelevant to our project.

The repository contains following locations for data:

• datasets
  • contains all available datasets in separate folders
• generalization/hierarchies
  • contains our defined generalization hierarchies per attribute and dataset
• results
  • all computed results (anonymised datasets, ML performance, etc.) are stored inside a folder structure inside results for each experiment
• paper_results
  • contains the results we used for analyses and plots in our paper
• figures
  • contains the figures used in our paper