Differential Privacy using K-Anonymity

To-do list

• Run 6 methods on 6 datasets
• Add L-diversity, Classic Mondrian (no hierarchies), Datafly algorithm
• Make NCP loss a separated module
• Implement DM, CAVG metrics
• Implement classification models (basic classifier, clustering)
• Run experiment on 6 datasets x 6 methods x 2 ML models
• Finish report
• (Improvement) T-closeness method, Incognito Algorithm
• (Optional) Simple Deanonymize Attack

Reports:

Report edit link:

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]

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