This is one of the most practical pieces of work in the differential privacy space since the original definition of DP. To have impact with data, it needs to be shared widely and easily combined (joined) with other data. For sensitive datasets, this has been hard. The setting either required trust or required interactivity with the publisher, limiting its spread. With this technique, one can publish data that can be distributed arbitrarily wide and far into the future without sacrificing privacy. The original publishers need not be in the picture. It also makes limited assumptions about the types of queries. I predict this will unlock a completely new set of products and applications for analyzing differential private data and create a new economy for data sharing.
1 comment
[ 2.9 ms ] story [ 15.9 ms ] thread