It's two, two nerddoms in one...

[jducoeur]

Okay, various of my friends may be amused by Google's page on Federated Learning.

On the one hand, this initiative is pretty intriguing: basically, a new approach to machine learning that specifically keeps all of the source data on user-owned devices, and encrypts everything being sent to the servers so that the Big Evil Company doing the machine-learning at least isn't pulling personal data into the cloud and potentially leaking it.

I'd be interested to hear opinions from folks who know this space better than I -- in principle the idea makes some sense, but I don't know enough to be able to validate whether it's as secure an approach as it claims.

Mind, it's clearly not a panacea: it's just a fix to one common privacy problem. And it's very self-interested on Google's part: since they control so many devices, they are in a better position to leverage this approach than most anybody else. Still, it seems better than the status quo.

What makes this much more amusing is that the linked page is not a dry description of the project. It's actually a description of the technology rendered as a comic book by Lucy Bellwood (a current favorite creator of mine) and Scott McCloud (one of the legends of the comics field). The project was apparently pretty amusing: as she says,

There were meetings and company policies and people used the word “synergy” in conversation without irony.

Anyway, the comic's worth reading at least if you have a clue about how this stuff works, and possibly even if not...

Comments

[verdantry]

Oh, Lucy Bellwood is so delightful! Thanks for linking this - it looks fascinating.

[dsrtao]

I'm thinking about the combination of parallel construction and Bloom filters.

Parallel construction: the methods we used to get this data were completely illegal, but now we have an answer and we can use the answer to go find a chain of evidence that we can use in court. So much easier!

Bloom filters: I'm not allowed to store this data permanently, but I can look at it in passing and can record a little metadata. A Bloom filter stores a factoid about a piece of data having been seen, being able to give definite No answers but merely high probabilities of Yes answers.


Oh, and "machine learning models are reifications of bias, known and unknown".

None of it is especially reassuring.

[drwex]

That's a neat and highly consumable thing. Reminds me of the various Cartoon Guide to XYZ books.

[jducoeur]

This sort of thing is kind of Lucy's forte -- she has a bunch of practice in teaching-through-comics. (Her particular passion is sailing on tall ships: her book [Baggywrinkles](https://lucybellwood.com/baggywrinkles/) is a delightful exploration of that field.)