[olliej]

My reading of the ORAM paper is that it is trying to solve a different problem - namely the situation where the threat you’re protecting against is an observer in the middle.

The purpose of differential privacy is to allow one a processor to compute an approximately accurate model without ever knowing the exact details of any origin.

Eg the example for ORAM is a secure element talking to main memory, and wanting to prevent an observer from knowing the actual structure of memory being used by the SE.

This is compared to the use case of differential privacy which is an endpoint wishing to process data without ever knowing what the actual source data was. For example as deployed by Apple: many metrics are bludgeoned with noise on the customers device before being sent to Apple servers. That way (theoretically) Apple never knows exact details of user behaviour, but can accumulate enough info to make approximations at scale.