Differential Privacy: How Apple Uses It to Learn About You Without Learning About YOU

Differential privacy is all set to be rolled out from iOS 10 onwards. The beta version of it was actually released at Apple’s Worldwide Developers’ Conference in June 2016.

What is differential privacy, anyway?

Apple says it is meant to collect data about a large group of users rather than individual users, to provide better services. Here, your personal data is collected, but ‘mathematical noise’ added to it to make it indistinguishable as YOUR data.

With Apple being Apple, there is no way anyone else is ever going to learn how this works, but one thing is for certain – your iOS 10 devices will be recording and sending more personal information than their predecessors. Otherwise, differential privacy, or DP, would be useless.

The idea is, adding mathematical noise to the collected information is expected to render it too obscure to be of real use to any one who may be interested in accessing that data (like a hacker). But it is also expected to have other repercussions. Here is an example.

Is differential privacy dangerous?

People who admire shoes offered for sale on shopping websites and then went on to read an online news article were startled to see an advertising banner for the same pair of shoes, being displayed on the news page they are reading. If this kind of personal information is being collected, what else are they recording?

Unfortunately, this trend is on the rise and more and more users are beginning to get uncomfortable with the idea of sharing where they have been, what they do online, who they talk to et al with third parties.

How NetFlix learned the importance of privacy the hard way

In 2007, Netflix ran a competition improve its recommendation algorithm.  As part of the contest, they released an actual sample of their users’ viewing data, some 480,000 of them. They took care to ensure that names were obscured and there was no way the users could be identified.

However, Arvind Narayanan and Vitaly Shmatikov of the University of Texas at Austin cross-referenced movie ratings, reviews and the dates on which they were submitted, against IMDB and showed that individual users could be identified from their data – proof that ‘anonymization’ doesn’t work. Their work led to a privacy suit against Netflix, which dropped its plans for a second competition in which they would have released even more data.

Apple’s DP is said to prevent this – but also aims to provide better QuickType and emoji recommendations for individual users. But many people haven’t been able to digest the fact that  Apple implemented it without testing and that they will never release their DP code.