Agenda Request - Learning inside a trusted server

[fhoering]

Agenda+: What do you want to discuss?

As a follow up of the presentation on Private Conversion Measurement via Global and Local DP (https://github.com/patcg/meetings/files/14936682/PATCG_Boston_private_learning.pdf) Criteo would like to present our working assumptions for learning inside a trusted server.

We will cover TEE based trusted servers and compare with MPC based approaches like the Hybrid Proposal/Firefox's Private Measurement API.

Covered topics

• assumed trust model
• experimental workflow
• constraints & experiments around differential privacy for models and offline metrics

Time

60 minutes including Q/A

Links

Executed benchmarks and sample code https://github.com/criteo-research/dp-sgd-ad-click-prediction

[AramZS]

@fhoering Hi, we're set up to have you present this during our TPAC meeting, but time is a little short, so can we aim for 30m for this presentation?

[fhoering]

@AramZS Hi, Yes. no problem. It can be done in 30 min also.

Do you think it is possible to move this item closer to the one from @mvono https://github.com/patcg/meetings/issues/192 in the timeline ? Both are very related as they represent different approaches with DP constraints. Normally we will do a common introduction and motivation section also.

We could maybe even say that both are part of a 60 min session.

[AramZS]

@fhoering Done. Updating the agenda now.

[AramZS]

@fhoering you and @mvono are one after the other in the 2:45 - 3:15 timeslot for @mvono and the 3:15 to 3:45 slot for you @fhoering. Thanks!