4 Local differential privacy for machine learning
This chapter covers
- Local Differential Privacy (LDP)
- Implementing the randomized response mechanism for LDP
- LDP mechanisms for one-dimensional data frequency estimation
- Implementing and experimenting with different LDP mechanisms for one-dimensional data
In the previous two chapters, we discussed centralized Differential Privacy, where there is a trusted data curator who collects data from individuals and applies different techniques to obtain differentially private statistics about the population. Then, the data curator publishes privacy-preserving statistics about this population. However, these techniques are unsuitable when individuals do not completely trust the data curator. Hence, various techniques to satisfy differential privacy in the local setting have been studied to eliminate the need for a trusted data curator. In this chapter, we will walk through the concept, the mechanisms, and the applications of the local setting of differential privacy, namely, local differential privacy.
This chapter will mainly look at how local differential privacy can be implemented in machine learning algorithms by looking at different examples and implementation codes. Toward the end of this chapter, we will also walk you through a case study on applying Locally Differentially Private Naïve Bayes Classification for real-world datasets.