Ted

Looking for a friendly, accessible way to learn about differential privacy? 🧐 Prefer simple explanations to ugly equations? 😇 Introducing my ✨ newly revamped ✨ blog post series about differential privacy, now with an intro page & table of contents 🌈 desfontain.es/privacy/friend…

Hahahaha I'm hearing that posts that mention other social networks by name are de-prioritized in "for you" timelines. What a loser move 🤣 Anyway I'm over there ⤵️

I now have a presence on Bluesky: bsky.app/profile/desfon… 🦋 If your main presence is on Bluesky, it'd be fantastic if you could follow bsky.app/profile/ap.bri…, so your posts are also mirrored on Mastodon, and I can follow you from there 💖

Super excited about this ✨ new blog post ✨, where I try to offer a super simple & intuitive explanation of the exponential mechanism, a fundamental building block in differential privacy 🧱 Feedback welcome! 💝 desfontain.es/privacy/choosi…

@ggharibi__ @Wikimedia @uscensusbureau Do you know of any publicly available information about this kind of deployment? I've heard some talk in academic circles but I don't know whether anything has been shipped to real products.

@TedOnPrivacy @Wikimedia @uscensusbureau Excellent list, but I was half-expecting to see LLMs and fine-tuning 😔

Updated my list of real-world deployments of differential privacy! 🚀 Changelog: - Added the @Wikimedia pageview release 🌐 - Added two releases by @uscensusbureau 🇺🇸 - Added details on a COVID-19 release from Google 💉 - Various maintenance fixes 🛠️ desfontain.es/privacy/real-w…

@PierreTholoniat @TumultLabs Yes indeed — I wrote the post long before this paper was on arXiv but it was published afterwards ^^ I'll try to find time to edit it next week to capture these results.

@TumultLabs @TedOnPrivacy This recent preprint (to appear at CCS '23) solves the first problem, no? arxiv.org/abs/2309.05901

Thanks a ton to @thegautamkamath, @ccanonne_, and @shortstein for their feedback and their help publishing this!

New, by me: what do we know about differential privacy composition theorems? What's missing, and would be useful to folks who are building practical DP software? 🔧 This blog post surveys existing results and raises a bunch of overlooked open problems. Feedback welcome! 💖

New blog post co-written by yours truly! In which we look at data clean rooms, what they do, what they often don't do, and why that can be a Problem™ 😬 Feedback super welcome as always! 🌈

first arxiv preprint ever🫣 “Publishing Wikipedia usage data with strong privacy guarantees,” which i’ll be presenting at TPDP 2023 in a few weeks arxiv.org/abs/2308.16298

Hear it at 2023 USENIX Conference on Privacy Engineering Practice and Respect: *Privacy-Preserving Analytics on the Ground* presented by Ryan Steed (Carnegie Mellon University). View the program and register now: bit.ly/pepr23prog #pepr23

The 2023 Caspar Bowden PET Award winner is: "The 2020 Census Disclosure Avoidance System TopDown Algorithm" #pets23

@HVitamines Thanks a ton for looking into this — I'll dive into your write-up this week and get back to you!

@TedOnPrivacy That's what I used in the section 3.2 of my answer to give an easy upper bound on this kl, and thus a (vi, xi) approx group wise zCDP where both vi and xi are both non increasing functions of x(D)_i

And now for something different: during a customer engagement, we stumbled upon a natural, but surprisingly tricky open problem in differential privacy. We couldn't figure it out, so we posted a simplified version of it as an open problem on DifferentialPrivacy.org 👀

@HVitamines Hi, thanks! To analyze the privacy guarantees of this mechanism under ρ-zCDP, one has to calculate the Rényi divergence of two Gaussian distributions of different variances. This is the hard part, unless we've missed something obvious?

@TedOnPrivacy I don't understand why this algorithm is not directly a solution to your problem.

@PET_Symposium @TumultLabs More info in the paper → arxiv.org/abs/2301.01998 🎓 And check out the full @PET_Symposium program; there it a ton of fantastic work in there: petsymposium.org/2023/program.p… ✨

If you're at @PET_Symposium next week, don't miss Marika Swanberg's talk about her research at @TumultLabs last summer: a novel DP mechanism for partition selection which, in contrast to prior work, scales horizontally 📈 In collaboration with Sam Haney & yours truly 😊

"When we want to protect individuals whose data is distributed across multiple rows, protecting individual rows is not enough." @TumultLabs introduces privacy IDs, .. "privacy identifiers enable the protection of individuals across large datasets, no matter how often they appear"

@const_charlie The mechanism is likely not optimal, but it still makes sense. The question is, though: can you prove a good per-group zCDP guarantee?

@TedOnPrivacy Lol it trivially kind of works if you're just measuring one statistic across all groups. Then the variance induced by the contribution of one group is proportional to the size of that group. Kind of silly though.

@const_charlie Thanks for the update! I had this on my list as things to check but didn't get around to it yet.

@TedOnPrivacy Actually this doesn't work! I just re-looked at the constraints and you want to output the k-ary histogram so a naive LDP mechanism would make every bin very noisy.

Many thanks to @thegautamkamath, @thejonullman and @shortstein for the helpful comments and the help with editing & publication ❤️