$\epsilon$-PrivateSMOTE

This project includes all the experimental evaluation of the $\epsilon$-PrivateSMOTE proposal.

Data

All data used including the original files and all transformed data variants.

• Original
• Transformed data with ARX tool.
• Data variants generated with deep learning, differentially private-based solutions and $\epsilon$-PrivateSMOTE.

Files description

code/tranformations Apply the transformations to the original data.

• task.py - puts in the queue all the original files to synthetise.

  Eg. python3 code/transformations/task_transf.py --input_folder "original" --type "PrivateSMOTE"

• main.py - apply transformation type: differentially private solutions (with Synthcity), deep learning solutions (with SDV) and $\epsilon$-PrivateSMOTE. It also includes the computational costs for generating each new data variant.

  Eg. python3 code/transformations/main.py --type "PrivateSMOTE"

code/modeling Machine Learning pipeline for predictive performance evaluation.

• task.py - puts in the queue all the files for modeling.

  Eg. python3 code/modeling/task.py --input_folder "output/oversampled/PrivateSMOTE"

• worker.py - calls apply_models (data handler) to run all the Machine Learning models.

  Eg. python3 code/modeling/worker.py --type "PrivateSMOTE" --input_folder "output/oversampled/PrivateSMOTE" --output_folder "output/modeling/PrivateSMOTE"

code/record_linkage Linkability for privacy risk evaluation.

• task_anonymeter.py - puts in the queue all the files for linkability analysis.

  Eg. python3 code/record_linkage/task_anonymeter.py --input_folder "output/oversampled/PrivateSMOTE"

• worker_anonymeter.py - calls anonymeter.

  Eg. python3 code/record_linkage/worker_anonymeter.py --type "PrivateSMOTE" --input_folder "output/oversampled/PrivateSMOTE" --output_folder "output/anonymeter/PrivateSMOTE"

code

• coverage.py - perform multiple generic data utility measures from SDVMetrics.

  Eg. python3 code/coverage.py --input_folder "PrivateSMOTE"

  Analysis in notebooks

• compt_costs.py - analysis of the computational costs in generating all data variants.

• results_coverage.py - analysis of the data utility measures for all data variants.

• results_performance.py - analysis of the predictive performance for all data variants.

• results_anonymeter.py - analysis of the linkability risk for all data variants.

• tradeoff.py - analysis of the tradeoff between predictive performance and linkability risk.

• bayesTest.py - analysis of the Bayesian Test for the oracle (best solutions in out of sample).

indexes.npy - indexes of the testing sets.

list_key_vars.csv - 5 random sets of quasi-identifiers for each data set.