Fedora considers “privacy-preserving” telemetry

Re:Fedora considers “privacy-preserving” telemetry by hedora on Friday July 07, 2023 @12:15PM (#36633957)

That's not true. Differential privacy just says that the output of the ML model will not change much if you add or remove a particular user, so you can't reverse the training process to infer what training data a given user provided.

It says nothing about whether or not you can join the output of multiple ML models with other telemetry to build a deanonymization model.

[ Parent | on HN ]

Re:Fedora considers “privacy-preserving” telemetry by bo1024 on Friday July 07, 2023 @07:32PM (#36639665)

Differential privacy satisfies a post-processing guarantee. It says that if you take the output of a differentially private process and do any amount of processing and combining with outside information, then you don't learn any more than you would have gotten with the outside information alone (up to epsilon).

[ Parent | on HN ]