3 Advanced concepts of differential privacy for machine learning

 

This chapter covers

  • Design principles of differentially private machine learning algorithms
  • Designing and implementing differentially private supervised learning algorithms
  • Designing and implementing differentially private unsupervised learning algorithms
  • Walking through designing and analyzing a differentially private machine learning algorithm

In the previous chapter we investigated the definition and general use of differential privacy (DP) and the properties of differential privacy that work under different scenarios (the postprocessing property, group property, and composition properties). We also looked into common and widely adopted DP mechanisms that have served as essential building blocks in various privacy-preserving algorithms and applications. This chapter will walk through how you can use those building blocks to design and implement multiple differentially private ML algorithms and how you can apply such algorithms in real-world scenarios.

3.1 Applying differential privacy in machine learning

In chapter 2 we investigated different DP mechanisms and their properties. This chapter will showcase how you can use those DP mechanisms to design and implement various differentially private ML algorithms.

3.1.1 Input perturbation

3.1.2 Algorithm perturbation

3.1.3 Output perturbation

3.1.4 Objective perturbation

3.2 Differentially private supervised learning algorithms

3.2.1 Differentially private naive Bayes classification

3.2.2 Differentially private logistic regression

3.2.3 Differentially private linear regression

3.3 Differentially private unsupervised learning algorithms

3.3.1 Differentially private k-means clustering

3.4 Case study: Differentially private principal component analysis

3.4.1 The privacy of PCA over horizontally partitioned data

sitemap