Hacker News Reader

Comment by nxzero | Hacker News Reader

Does anyone know if the "differential privacy" is done on the device before send the data to Apple? Seems unlikely, but oddly couldn't get an answer from Apple.

View on HN

From the article: "In the version of differential privacy Apple is using, known as the local model, software on a person’s device adds noise to data before it is transmitted to Apple. The company never gets hold of the raw data."

Yes, that's the whole point. Apple's servers never see the raw data.

It looks like it is done on the device based on the framework header : https://github.com/JaviSoto/iOS10-Runtime-Headers/tree/maste...

Interesting. I thought that with DP the user data is anonymized by a random function with the value range based on the complete sample set. I also expect that the data is encrypted before transmission and I expect some kind of unique signature which is somehow stored on the device. The transmission itself is certainty logged on client and server side.

If I understand this framework correct the epsilon range for the randomization can be selected by hand before the transmission and also the number of collected parameters is not limited. Why is this called "differential privacy"?

Perhaps named because the system is based on differences between the true value and the reported value .... and then summing the aggregate data removes these 'differences'.

If a data collector can choose the epsilon range and the linked parameters by oneself, s/he could choose small (irrelevant) or too large (expandable) ranges and could try to cross reference the result via data-mining (already collected data).

True.

"[...] software on a person’s device adds noise to data before it is transmitted to Apple. The company never gets hold of the raw data."

So I guess yes?

Why unlikely? Doing it anywhere but in the device would defeat the purpose of the whole system.

Reply on news.ycombinator.com