4 comments

[ 32.8 ms ] story [ 228 ms ] thread
While this works for some datasets with known statistical properties, it seems unlikely to become a universal option. Differential Privacy relies on the intuition that something is only private of it happens once in a dataset. Multiple times? Must be something humans have in common.

The issue is that you need massive datasets for everything non-personal to start repeating. And one of the parallel constraints of protecting privacy is not centralising massive datasets.

Encrypted computation needed.

Indeed, Differential Privacy is very useful, and it's an incredible step in the right direction for some problems, but it's no silver bullet. Even Homomorphic database encryption is only useful for some use cases.

Ultimately there is no escaping understanding the statistical consequences of data.

Encrypted computation does not protect privacy. You can still compute arbitrary, privacy-violating functions
I have mixed feelings on synthetic data. We don't honestly know what synthetic data is missing when we use it to flush out or replace more privileged or classified datasets.