Data anonymization using k-Anonymity

✔️ Team's contribution (Amir, Nikita, Karol)

Repository was forked from here: https://github.com/kaylode/k-anonymity

Our contribution:

1. We prepared our own dataset to be anonymized, together with quasi-identifiers' hierarchies definition. All can be found in /data/covid
2. We anonymized the dataset using the datafly algorithm for different k values. Results can be found in /results/covid
3. We calculated CAVG, DM and NCP metrics for different k-values. Results are stored in metrics_calculated.csv
4. We generated plots with above-mentioned metrics, which can be found in /plots

• Anonymization of the data, together with metrics calculation and saving them into csv can be called this way:

python metrics_calculate.py

Variables methods, dataset, k_array inside metrics_calculate.py are responsible, respectively, for the choice of the algorithm, data and k values. By default covid dataset and datafly algorithm are being used. The script might take a while to run.

• Plots can then be generated by calling:

python metrics_plot.py

Explanations below come directly from the forked repository and have not been changed

✔️ Experiments

• Provides 5 k-anonymization method:
  • Datafly
  • Incognito
  • Topdown Greedy
  • Classic Mondrian
  • Basic Mondrian
• Implements 3 anonymization metrics:
  • Equivalent Class size metric (CAVG)
  • Discernibility Metric (DM)
  • Normalized Certainty Penalty (NCP)
• Implements 3 classification models:
  • Random Forests
  • Support Vector Machines
  • K-Nearest Neighbors

📖 Reports

• Report edit link: link
• Slide link: link

Folder Structure

• A dataset must comes with a .csv file contains features information and a hierarchy folder which contains predefined generalization hierarchies for its QID attributes.

this repo
│   anonymize.py
|
└───data  
│   │
│   └───adult
│       │   adult.csv
│       └───hierarchies
│       │     adult_hierarchy_workclass.csv
│       │     ....

• Here is an example for a generalization hierarchy of the 'workclass' attribute from ADULT dataset, described in adult_hierarchy_workclass.csv, which is a csv file using ";" as delimiter

Private;Non-Government;*
Self-emp-not-inc;Non-Government;*
Self-emp-inc;Non-Government;*
Federal-gov;Government;*
Local-gov;Government;*
State-gov;Government;*
Without-pay;Unemployed;*
Never-worked;Unemployed;*

which describes this tree:

---

🌟 Executing

To anonymize dataset, run:

python anonymize.py --method=<model_type> --k=<k-anonymity> --dataset=<dataset_name>

• model_type: [mondrian | classic_mondrian | mondrian_ldiv | topdown | cluster | datafly]
• dataset_name: [adult | cahousing | cmc | mgm | informs | italia]

Results will be in results/{dataset}/{method} folder

To run evaluation metrics on every combination of algorithms, datasets and value k, run:

python visualize.py

Results will be in demo/{metrics.png, metrics_ml.png}

K-Anonymity examples

Before anonymization	After anonymization with k = 2

Evaluation Metrics

Evaluate anonymization using information loss metrics

Evaluate anonymization using classification models

References:

• Basic Mondrian, Top-Down Greedy, Cluster-based (https://github.com/fhstp/k-AnonML)
• L-Diversity (https://github.com/Nuclearstar/K-Anonymity, https://github.com/qiyuangong/Mondrian_L_Diversity)
• Classic Mondrian (https://github.com/qiyuangong/Mondrian)
• Datafly Algorithm (https://github.com/nazilkbahar/python-datafly)
• Normalized Certainty Penalty from Utility-Based Anonymization for Privacy Preservation with Less Information Loss
• Discernibility, Average Equivalent Class Size from A Systematic Comparison and Evaluation of k-Anonymization Algorithms for Practitioners
• Privacy in a Mobile-Social World
• Code and idea based on k-Anonymity in Practice: How Generalisation and Suppression Affect Machine Learning Classifiers