Privacy preserving single post analytics

Ryan Rogers, Subbu Subramaniam, Lin Xu12 Dec 2023

Problem

LinkedIn shows a post's author the demographics of viewers: top companies, job titles, locations, industries and company sizes. Viewer identities are hidden, but the breakdowns update in real time. An author who watches them change after a new view can deduce the company, title and location of the member who just viewed. Three attributes (company, job title, location) uniquely identify more than a third of weekly active members. With top-20 breakdowns about 9% of initial viewers on sampled posts were identifiable; top-5 cut this to under 2%, still too high.

Example from the article: a running counter of Data Scientist viewers goes 1,1,1,1,2,2,2,2 over eight viewers, so every change in it reveals that the latest viewer was a Data Scientist.

Solution in brief

Add differentially private noise to the breakdowns, but use correlated noise from the binary mechanism for continual observation, so the privacy loss grows only logarithmically with the number of viewers. A mid-tier service (PEDAL) sits between the application and the OLAP store, picks the algorithm per query, and the team verifies the result with an empirical re-identification metric.

Questions this case answers
  • How to publish frequently updated top-k group counts over a stream so that the reader cannot infer a single individual's attributes from changes between updates
  • How to empirically measure re-identification risk of released aggregates and balance it against accuracy of the top results
  • How to show live demographic breakdowns of an audience without letting the viewer of the report deduce who a single member is?
  • How to add differential privacy to a real-time dashboard that updates after every new event without the privacy budget exploding?

Tools and methods

  • Continual-observation differential privacy (binary mechanism) — noisy running counts with privacy loss logarithmic in stream length
  • Re-identification risk metric — replay analytics on sampled posts and count identifiable viewers
  • Apache Pinot — real-time OLAP store that serves the true counts
Continue reading — it's free

Google or a one-time email code. No password.