Differential Privacy Synthetic PII is a set of Python algorithms that enable generation of differential privacy, fully synthetic personally identifiable information (PII) data and then impute additional insights based on real data. The algorithms enable generation of data which can be distributed easily without revealing private information. A core piece behind our differential privacy data is the introduction of randomness behind the imputation of certain features while not impacting the overall dataset's valid inferences. By introducing randomness to the data generation, differential privacy can create deniability behind identifiable features of individuals.
- Python (>= 3.8.5)
- NumPy (1.17.0)
- SciPy
- Dython
- MLxtend
- Faker
- Pandas
- gender_guesser
- re
- tqdm
- datetime
- random
- matplotlib
A. Create a bulk data request CSV file with the following format (keep column names the same):
| Country | Year | Records |
|---|---|---|
| RUSSIA | 2016 | 457 |
| MEXICO | 2020 | 1102 |
| CHINA | 2018 | 632 |
B. Update line 9 if needed in the script fake_pii_generator.py to point to the file made in step A. Then execute script.
C. Clean the differential privacy synthetic PII with the fake_data_clean.py script. This script primarily removes any person titles (Dr., Ing., Lic., etc.), cleans gender categories, and removes any records with P.O. Box or military addresses.
Due to the introduction of randomness behind many of the PII features, there is a chance the dataset may not have much valid inferences. Therefore, adjusting some of the features such as Age or other dates, can create custom context and insights. The script age_update.py is an example of imputed insights that was adjusted on a real distribution dataset by the U.S. Immigration Customs Enforcement Agency (ICE) made available to the public here.