Differential Privacy

What is Differential Privacy

Differential Privacy is a relatively new filed, that was applied to statistical queries in 2002. And eventually applied to machine learning.

The general goal of differential privacy is that different kinds of statistical analysis does not compromise the privacy. Meaning that if we have a database containing information about certain users, we don’t want any kind of analysis or queries, to compromise the privacy or data for a particular individuals. We are trying to learn something about the dataset and not about specific individuals.

Differential Privacy Definition

Anything that can be learned about a participant from the statistical database, can be learned without access to the database.

“Differential privacy describes a promise made by a data holder, to a data subject, and the promise is like that” you will not be affected, adversely or otherwise by allowing your data to be used in a study or analysis, no matter what other studies, data sets or information sources are available”

-Cynthia Dwork, Algorithmic foundations

Can’t we just Anonymize the data?

At first this may seem like a good idea, since we won’t know which row of the data belongs to which user. But what happens if another dataset with the same users is released somewhere else. In that case by comparing and studying both datasets, we would be able to de-anonymize a large number of the users.

Considering Cynthia’s robust definition for differential privacy. Any source of information is an important constraint. And that’s why simple anonymization is not strong enough for achieving privacy.

So, how can we better achieve Differential Privacy?

I am writing this article in part with the Udacity’s secure and private AI Scholarship Challenge, as a way to share what I have learned so far.

In the upcoming articles, I’ll take you along my learning journey for achieving differential privacy.

#60daysofudacity #secureandprivateai