Building Machine Learning Pipelines

Chapter 14. Data Privacy for Machine Learning

In this chapter, we introduce some aspects of data privacy as they apply to machine learning pipelines. Privacy-preserving machine learning is a very active area of research that is just beginning to be incorporated into TensorFlow and other frameworks. We’ll explain some of the principles behind the most promising techniques at the time of writing and show some practical examples for how they can fit into a machine learning pipeline.

We’ll cover three main methods for privacy-preserving machine learning in this chapter: differential privacy, federated learning, and encrypted machine learning.

Data Privacy Issues

Data privacy is all about trust and limiting the exposure of data that people would prefer to keep private. There are many different methods for privacy-preserving machine learning, and in order to choose between them, you should try to answer the following questions:

• Who are you trying to keep the data private from?

• Which parts of the system can be private, and which can be exposed to the world?

• Who are the trusted parties that can view the data?

The answers to these questions will help you decide which of the methods described in this chapter best fits your use case.

Why Do We Care About Data Privacy?

Data privacy is becoming an important part of machine learning projects. There are many legal requirements surrounding user privacy, such as the EU’s General Data Protection Regulation (GDPR), which went into effect ...