Ryan retweeted
Another result of Durfee and @RowRyRo that should be better known. They showed that the exponential mechanism in DP has better composition properties than generic DP algorithms. One application: Selecting the top k elements of some set. It turns out you can do it in one shot.
One-shot DP Top-k mechanisms
ift.tt/3sglVpD
Check out this post with David Durfee on how iteratively applying the exponential mechanism can be done in one shot with Gumbel noise.
One-shot DP Top-k mechanisms
ift.tt/3sglVpD
This is a sequel to this other post about the exponential mechanism: xitter.cf/RowRyRo/status/1414626…
Check out this latest post with @shortstein on the improved privacy analysis of the exponential mechanism. differentialprivacy.org/expo…
Ryan retweeted
Originating from conversations at the 2020 Google Workshop on Federated Learning and Analytics (shorturl.at/myAY8), 26 @GoogleAI researchers and 27 academic researchers began a year-long collaboration resulting in A Field Guide to Federated Optimization
Ryan retweeted
An important result that has gone under-appreciated for a few years: the exponential mechanism is about twice as private as you think (at least under composition).
Cool new post by @shortstein and @RowRyRo showing how to save a factor of 2 in the privacy analyss of the exponential mechanism! Fancy new algorithms are great, but improving a ubiquitous tool with no downsides is hard to beat!
differentialprivacy.org/expo…
Check out this latest post with @shortstein on the improved privacy analysis of the exponential mechanism. differentialprivacy.org/expo…
In follow up work with David and @JinshuoD, we studied optimal composition of BR mechanisms. Interestingly, we found there is a difference between the privacy loss bound when the mechanisms are chosen adaptively vs non-adaptively. arxiv.org/abs/1909.13830
Continuing this line of work with Mark Cesar, we then studied (optimal) privacy loss bounds when your DP system consists of BR (exponential mechanisms), pure DP (Laplace mechanisms), and CDP (Gaussian mechanisms). arxiv.org/abs/2004.07223
Check out our fearless leader in Applied Research at @LinkedInEng talk about the great work from our team, including differential privacy!
Today we’re joined by Parvez Ahammad of @LinkedIn @LinkedInEng to discuss how to identify products that are useful across multiple lines of business, quantitative methodologies to identify unintended consequences in experimentation, GreyKite and much twimlai.com/agile-applied-ai…
Ryan retweeted
CfP for the workshop on Theory and Practice of Differential Privacy (TPDP 2021, co-located with @icmlconf #ICML2021) is up! Deadline: May 28. Submit work on any aspects of DP! Co-chaired with @radcummings. Invited speakers are Katrina Ligett and @RowRyRo. tpdp.journalprivacyconfident…
Ryan retweeted
Did you miss TPDP 2020? If you were unable to attend, or want to watch again, videos of the talks are available here: tpdp.journalprivacyconfident…
Ryan retweeted
The **Algorithms** group @MSFTResearch Redmond is hiring! It is the sister group to our ML Foundations group, and has already strong presence in differential privacy and coding!
Post Doc: careers.microsoft.com/us/en/…
FTE: careers.microsoft.com/us/en/…
Nice connection between exponential noise and Permute-and-Flip!
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
Ryan retweeted
📺🎙 Today’s privacy_infra() #PrivacyEng event is about to begin (10am PT)! Join us at zoom.us/webinar/register/801…
Excited to present tomorrow at the virtual privacy engineering meetup–hosted by @transcend_io! @LinkedInEng
📣We're thrilled to add LinkedIn's Ryan Rogers and the Wikimedia Foundation's @pantojacoder to the speaker list for our next privacy_infra() #privacyeng meetup on 11/12. Register: zoom.us/webinar/register/481…
Ryan retweeted
🗓We've got our last #privacyinfra event for 2020 this Thursday at 10am PT, with privacy eng. talks from LinkedIn, Wikipedia, and more. #privacyeng Register here: zoom.us/webinar/register/481…