Hacker News Reader

Comment by z77dj3kl | Hacker News Reader

There is a whole field (now relatively mainstream) of differential privacy, concerned with answering questions such as "can I be correlated and de-anonymized across queries" (query might be "what's your current cohort id?").

Is FLoC not built on sound principles of differential privacy? That would be a big shame on Google.

EDIT: Huge shame on Google! From their FLoC whitepaper: "We want to emphasize that, even though differential privacy is now the de facto privacy notion in industry and academia, we decided against using it as our privacy measure for building audiences."

What in the world are they thinking?!

View on HN

Differential privacy is useful for training or updating a public model where individuals' features should be kept private.

In floc's case the model is public but isn't being trained on individual's features in realtime, only used for inference as far as the proposal says, e.g. the proof of concept stage will develop a fixed model that all browser instances (of a given vendor) share. Individuals' features are kept private to the extent that the model output can't be effectively reverse-engineered.

Differential privacy probably also won't be useful in the POC stage because the training will require accurate labels which defeats privacy.

differential privacy is good for answering population questions like "how many people in my dataset have property x?". it's a lot less clear how to apply it to something as granular as serving personalized ads. and as the example demonstrates, this compounds if you're doing it repeatedly with data that keeps getting updated. to the best of my knowledge, "differentially private personalized ads" is a hard problem, and maybe just a contradiction in terms.

I think it’s Google’s responsibility to make it clear, though, either by putting in the theoretical work to apply differential privacy or proposing a refinement of the concept that allows them to. It’s like those people who propose grand new theories of physics without using any math; if you can’t connect your ideas to what’s come before, people will be rightfully suspicious whether they’re built on quicksand.

Reply on news.ycombinator.com