The true promise of Differential Privacy is NOT - "taking your experiment and making it private"

The true promise of DP is a new social institution wherein your model now trains on data from 1000s of hospitals instead of just one

False comparisons are where DP gets a bad rap
If you compare "training with DP" and then training *on the same dataset* "without DP" and go "awww man - DP is terrible!" ... this is the wrong comparison.

A true 1:1 comparison is a *world* with/without DP infrastructure.
(and to be clear - we still live in a world without the kind of infrastructure I'm talking about)
You can follow @iamtrask.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: