PETWorks-framework

A framework for validating PET enhanced data.

Data Privacy is the keystone promoting stronger and deeper analysis of problem in our society. PET (Privacy Enhancing Tool) is a great helper for making all these happened. However, a trust-worthy and easy-to-use validation tool for PET is still rare.

Here we provide a framework dealing with the validation problem of PET enhanced data.

Showcase

Validation of Data Processed With Federated Learning

from PETWorks import dataProcess, PETValidation, report

gradient = "/home/Doc/gradient"
model = "/home/Doc/model"
originalData = "/home/Doc/o.png"

recoveredData = dataProcess(model, gradient, "FL", "recover")
result = PETValidation(recoveredData, originalData, "FL")
report(result, "web")

Measurement of the Re-Identification Risk

from PETWorks import PETValidation, report

originalData = "data/deidentifiedData.csv"

result = PETValidation(originalData, None, "ReidentificationRisk")
report(result, "json")

Execution Result

{
    "Re-identification Risk": 1.0
}

Measurement of the Ambiguity Metric

from PETWorks import PETValidation, report
from PETWorks.attributetypes import QUASI_IDENTIFIER, SENSITIVE_ATTRIBUTE

originalData = "data/adult.csv"
anonymizedData = "data/adult_anonymized.csv"
dataHierarchy = "data/adult_hierarchy"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "education": QUASI_IDENTIFIER,
    "marital-status": QUASI_IDENTIFIER,
    "native-country": QUASI_IDENTIFIER,
    "occupation": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
    "salary-class": QUASI_IDENTIFIER,
    "sex": QUASI_IDENTIFIER,
    "workclass": SENSITIVE_ATTRIBUTE,
}

result = PETValidation(
    originalData, anonymizedData,
    "Ambiguity",
    dataHierarchy=dataHierarchy,
    attributeTypes=attributeTypes
)
report(result, "json")

Execution Result

$ python3 ambiguity.py
{
    "ambiguity": 0.72714009672634
}

Measurement of the Precision Metric

from PETWorks import PETValidation, report
from PETWorks.attributetypes import QUASI_IDENTIFIER, SENSITIVE_ATTRIBUTE

originalData = "data/adult.csv"
anonymizedData = "data/adult_anonymized.csv"
dataHierarchy = "data/adult_hierarchy"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "education": QUASI_IDENTIFIER,
    "marital-status": QUASI_IDENTIFIER,
    "native-country": QUASI_IDENTIFIER,
    "occupation": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
    "salary-class": QUASI_IDENTIFIER,
    "sex": QUASI_IDENTIFIER,
    "workclass": SENSITIVE_ATTRIBUTE,
}

result = PETValidation(
    originalData, anonymizedData,
    "Precision",
    dataHierarchy=dataHierarchy,
    attributeTypes=attributeTypes
)
report(result, "json")

Execution Result

$ python3 precision.py
{
    "precision": 0.7271401100722763
}

Measurement of the Non-Uniform Entropy Metric

from PETWorks import PETValidation, report
from PETWorks.attributetypes import QUASI_IDENTIFIER, SENSITIVE_ATTRIBUTE

originalData = "data/adult.csv"
anonymizedData = "data/adult_anonymized.csv"
dataHierarchy = "data/adult_hierarchy"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "education": QUASI_IDENTIFIER,
    "marital-status": QUASI_IDENTIFIER,
    "native-country": QUASI_IDENTIFIER,
    "occupation": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
    "salary-class": QUASI_IDENTIFIER,
    "sex": QUASI_IDENTIFIER,
    "workclass": SENSITIVE_ATTRIBUTE,
}

result = PETValidation(
    originalData, anonymizedData,
    "Non-Uniform Entropy",
    dataHierarchy=dataHierarchy,
    attributeTypes=attributeTypes
)
report(result, "json")

Execution Result

$ python nonUniformEntropy.py
{
    "Non-Uniform Entropy": 0.6740002378300514
}

Measurement of the AECS Metric

from PETWorks import PETValidation, report
from PETWorks.attributetypes import QUASI_IDENTIFIER, SENSITIVE_ATTRIBUTE

originalData = "data/adult.csv"
anonymizedData = "data/adult_anonymized.csv"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "education": QUASI_IDENTIFIER,
    "marital-status": QUASI_IDENTIFIER,
    "native-country": QUASI_IDENTIFIER,
    "occupation": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
    "salary-class": QUASI_IDENTIFIER,
    "sex": QUASI_IDENTIFIER,
    "workclass": SENSITIVE_ATTRIBUTE,
}

result = PETValidation(
    originalData, anonymizedData,
    "AECS",
    attributeTypes=attributeTypes
)
report(result, "json")

Execution Result

$ python aecs.py
{
    "AECS": 0.9992707253704929
}

Compute the k-anonymity

from PETWorks import PETValidation, report
from PETWorks.attributetypes import QUASI_IDENTIFIER

anonymizedData = "data/adult_anonymized.csv"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "sex": QUASI_IDENTIFIER,
}

result = PETValidation(
        None, anonymizedData, "k-anonymity", attributeTypes=attributeTypes, k=6
)
report(result, "json")

Execution Result

$ python3 k-anonymity.py
{
    "k": 6,
    "k-anonymity": true
}

Compute the δ-presence

from PETWorks import PETValidation, report
from PETWorks.attributetypes import SENSITIVE_ATTRIBUTE, QUASI_IDENTIFIER


origin = "data/delta.csv"
anonymized = "data/delta_anonymized.csv"
dataHierarchy = "data/delta_hierarchy"

attributeTypes = {
    "zip": QUASI_IDENTIFIER,
    "age": QUASI_IDENTIFIER,
    "nationality": QUASI_IDENTIFIER,
    "salary-class": SENSITIVE_ATTRIBUTE
}

result = PETValidation(
        origin, anonymized, "d-presence", dataHierarchy=dataHierarchy, attributeTypes=attributeTypes, dMin=1/2, dMax=2/3
    )
report(result, "json")

Execution Result

$ python3 d-presence.py
{
    "dMin": 0.5,
    "dMax": 0.6666666666666666,
    "d-presence": true
}

Compute the Profitability

from PETWorks import PETValidation, report
from PETWorks.attributetypes import QUASI_IDENTIFIER, INSENSITIVE_ATTRIBUTE

origin = "data/delta.csv"
anonymized = "data/delta_anonymized.csv"
dataHierarchy = "data/delta_hierarchy"

attributeTypes = {
    "zip": QUASI_IDENTIFIER,
    "age": QUASI_IDENTIFIER,
    "nationality": QUASI_IDENTIFIER,
    "salary-class": INSENSITIVE_ATTRIBUTE
}

result = PETValidation(
    origin,
    anonymized,
    "profitability",
    dataHierarchy=dataHierarchy,
    attributeTypes=attributeTypes,
    allowAttack=True,
    adversaryCost=4,
    adversaryGain=300,
    publisherLost=300,
    publisherBenefit=1200
)
report(result, "json")

Execution Result

$ python3 profitability.py
{
    "allow attack": true,
    "adversary's cost": 4,
    "adversary's gain": 300,
    "publisher's loss": 300,
    "publisher's benefit": 1200,
    "isProfitable": true
}

Compute the l-diversity

from PETWorks import PETValidation, report
from PETWorks.attributetypes import SENSITIVE_ATTRIBUTE, QUASI_IDENTIFIER

anonymized = "data/inpatient_anonymized.csv"

attributeTypes = {
    "zipcode": QUASI_IDENTIFIER,
    "age": QUASI_IDENTIFIER,
    "nationality": QUASI_IDENTIFIER,
    "condition": SENSITIVE_ATTRIBUTE
}

result = PETValidation(
    None, anonymized, "l-diversity", attributeTypes=attributeTypes, l=3
)
report(result, "json")

Execution Result

$ python3 l-diversity.py
{
    "l": 3,
    "fulfill l-diversity": true
}

Compute the t-closeness

from PETWorks import PETValidation, report
from PETWorks.attributetypes import (
    SENSITIVE_ATTRIBUTE,
    QUASI_IDENTIFIER,
)

anonymized = "data/patient_anonymized.csv"
dataHierarchy = "data/patient_hierarchy"

attributeTypes = {
    "ZIPCode": QUASI_IDENTIFIER,
    "Age": QUASI_IDENTIFIER,
    "Disease": SENSITIVE_ATTRIBUTE,
}

result = PETValidation(
    None,
    anonymized,
    "t-closeness",
    dataHierarchy=dataHierarchy,
    attributeTypes=attributeTypes,
    tLimit=0.376,
)
report(result, "json")

Execution Result

$ python3 t-closeness.py
{
    "t": 0.376,
    "fulfill t-closeness": true
}

Anonymize with the k-anonymity

from PETWorks import PETAnonymization, output
from PETWorks.attributetypes import *

originalData = "data/adult.csv"
dataHierarchy = "data/adult_hierarchy"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "sex": QUASI_IDENTIFIER,
}

result = PETAnonymization(
    originalData,
    "k-anonymity",
    dataHierarchy,
    attributeTypes,
    maxSuppressionRate=0.6,
    k=6,
)

output(result, "output.csv")

Anonymize with the δ-presence

from PETWorks import PETAnonymization, output
from PETWorks.attributetypes import *

originalData = "data/adult.csv"
dataHierarchy = "data/adult_hierarchy"
subsetData = "data/adult10.csv"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
}

result = PETAnonymization(
    originalData,
    "d-presence",
    dataHierarchy,
    attributeTypes,
    maxSuppressionRate=0.6,
    dMin=0.0,
    dMax=0.8,
    subsetData=subsetData,
)

output(result, "output.csv")

Anonymize with the l-diversity

from PETWorks import PETAnonymization, output
from PETWorks.attributetypes import *

originalData = "data/adult.csv"
dataHierarchy = "data/adult_hierarchy"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
    "workclass": SENSITIVE_ATTRIBUTE,
}

result = PETAnonymization(
    originalData,
    "l-diversity",
    dataHierarchy,
    attributeTypes,
    maxSuppressionRate=0.6,
    l=6,
)

output(result, "output.csv")

Anonymize with the t-closeness

from PETWorks import PETAnonymization, output
from PETWorks.attributetypes import *

originalData = "data/adult.csv"
dataHierarchy = "data/adult_hierarchy"

attributeTypes = {
    "age": QUASI_IDENTIFIER,
    "race": QUASI_IDENTIFIER,
    "workclass": SENSITIVE_ATTRIBUTE,
}

result = PETAnonymization(
    originalData,
    "t-closeness",
    dataHierarchy,
    attributeTypes,
    maxSuppressionRate=0.6,
    t=6,
)

output(result, "output.csv")

How it works?

Module	Description
PET Enhanced Data Process	This module aims to load and process (e.g. recover) PET-enhanced data accordingly to the data and the PET it used.
PET Validation	This module has validation methodologies built in. Returning the result of the validation.
Report	This module handles the output format of the report. It could be text-based on the terminal or GUI-based showing on the web browser.

Web Report

Here is the showcase of the web report.

The web report also shows the process of recovery.

Current Status

This project is now maintained by Telecom Technology Center, Taiwan. We're now providing just a really simple showcase demonstrating how we can use this framework to validate PET protection of the data using the technology of federated learning. We still need plenty of implementation for every module. This is a very early stage project. Stay tuned!