Bouw een differentieel privéklaasificador

In deze oefening bouw en train je een private Gaussian Naive Bayes-classifier op de Penguin-gegevensset om te classificeren of een pinguïn mannelijk of vrouwelijk is.

K-anonimiteit werkt niet goed met hoog-dimensionale of diverse gegevenssets vanwege aanzienlijke theoretische en empirische beperkingen, de “vloek van dimensionaliteit”. Naarmate het aantal features of dimensies toeneemt, groeit de hoeveelheid data die we nodig hebben om goed te generaliseren exponentieel. Dat is een van de redenen waarom differential privacy het huidige, voorkeurs-privacymodel is. Epsilon is onafhankelijk van achtergrondkennis en “begrensd” de gevoelige informatie.

De DataFrame is geladen als penguin_df en gesplitst in X_train, y_train, X_test en y_test. De private modelklasse is geïmporteerd als dp_GaussianNB.

Deze oefening maakt deel uit van de cursus

Dataprivacy en anonimisering in Python

Cursus bekijken

Oefeninstructies

Maak een dp_GaussianNB-classifier zonder parameters.
Fit het eerder gemaakte model op de data zonder parameters.
Bereken de score van het private model op basis van de testdata.

Praktische interactieve oefening

Probeer deze oefening eens door deze voorbeeldcode in te vullen.

# Built the private classifier without parameters
dp_clf = ____

# Fit the model to the data
____(X_train, y_train)

# Print the accuracy score
print("The accuracy with default settings is ", ____(X_test, y_test))

Code bewerken en uitvoeren

Deze oefening maakt deel uit van de cursus

Dataprivacy en anonimisering in Python

SkillTag.level.advancedSkillTag.label

4.9+

Begin de cursus gratis

Get ready to apply anonymization techniques such as data suppression, masking, synthetic data generation, and generalization. In this chapter, you’ll learn how to distinguish between sensitive and non-sensitive personally identifiable information (PII), quasi-identifiers, and the basics of the GDPR. You'll also encounter real-life examples of what can go wrong if you don't follow these best practices.

Exercise 1: What's private, and why do we care?Exercise 2: Privacy is power Exercise 3: Is it sensitive or non-sensitive?Exercise 4: Suppression of sensitive attributes Exercise 5: Data masking and data generation with Faker Exercise 6: Masking sensitive PII Exercise 7: Removing names with faker Exercise 8: Anonymizing with data generalization Exercise 9: Reducing identification risk with generalization Exercise 10: Data aggregation and data generalization Exercise 11: Top and bottom coding White House salaries

Discover how to anonymize data by sampling from datasets following the probability distribution of the columns. You’ll then learn how to apply the k-anonymity privacy model to prevent linkage or re-identification attacks and use hierarchies to perform data generalization in categorical variables.

Exercise 1: Anonymizing categorical data Exercise 2: Explore the distribution of data Exercise 3: Sampling from the same probability distribution Exercise 4: Anonymizing continuous data Exercise 5: Different distributions Exercise 6: Sampling from the best continuous distribution Exercise 7: Introduction to K-anonymity Exercise 8: Privacy attributes Exercise 9: Generalizing into ranges Exercise 10: Generalizing data using hierarchies Exercise 11: Using hierarchies for categorical data Exercise 12: K-anonymizing a dataset

Learn about differential privacy, the model used by major technology companies such as Apple, Google, and Uber. In this chapter, you’ll explore data by generating private histograms and computing private averages in data. You’ll also create differentially private machine learning models that allow businesses to increase the utility of their data.

Exercise 1: Introductie tot differential privacy Exercise 2: Epsilon (ϵ): het magische getal Exercise 3: Histogrammen met differential privacy Exercise 4: Privacybudgetten Exercise 5: Privacybudgetten gebruiken Exercise 6: Wanneer er geen budget meer is Exercise 7: Data verkennen met een privacybudget-accountant Exercise 8: Differentiële private Machine Learning-modellen Exercise 9: Bouw een differentieel privéklaasificador

Huidige oefening

Exercise 10: Salarissen voorspellen Exercise 11: Differentiële private clusteringmodellen Exercise 12: Preprocessing van data Exercise 13: Klanten segmenteren

In this final chapter, you’ll learn how to apply dimensionality reduction methods such as principal component analysis (PCA) to anonymize large multi-column datasets. You’ll then use Faker to generate realistic and consistent datasets, and scikit-learn to create synthetic datasets that follow a normal distribution. Lastly, you’ll tie everything you learned in this course together as you combine multiple techniques to safely release datasets to the public.

Exercise 1: PCA for anonymization Exercise 2: Anonymization of high-dimensional data Exercise 3: Data masking with PCA Exercise 4: Generating realistic datasets with Faker Exercise 5: Consistent synthetic dataset Exercise 6: Datasets with the same probabilistic distribution Exercise 7: Creating synthetic datasets using scikit-learn Exercise 8: Generating datasets for classification Exercise 9: Generating datasets for clustering Exercise 10: Safely release datasets to the public Exercise 11: Exploring and pseudonymizing a dataset Exercise 12: Preparing employee data for safe release Exercise 13: Great work!