Hands-On Differential Privacy

Chapter 8. Differentially Private Statistical Modeling

Statistical modeling encompasses a broad gamut of tools used to find patterns in data and make predictions. Statistical models fitted with a differentially private algorithm protect the privacy of individuals who are in the data used to train the model. This chapter focuses on introducing you to differentially private algorithms that release hyperparameters. These algorithms are executed when you are fitting DP statistical models.

It is important not to miss the distinction that a differentially private model is simply a model whose hyperparameters are differentially private releases made on the training data. Therefore, inferences made on DP models are not private. At best, the model inference may qualify as a stable data-set-to-data-set transformation. The purpose of DP fitting is to make the model itself resilient against attacks intended to violate the privacy of individuals in the training data.

This chapter covers the following topics:

• Private inference

• Private linear regression

• Algorithm selection

• Private naive Bayes

• Private decision trees

• The relationship between model fitting, model parameters, and potential privacy violations

This chapter assumes a working knowledge of statistical modeling as a prerequisite in order to maintain focus on differential privacy.

Private Inference

Inference refers to feeding a data set to a trained model to obtain a transformed data set. There are many contexts where inference ...