You can find a detailed explanation and the current state of research and implementation here as well as our WIP paper. What follows are intros to the two major building blocks.
Recent advances in Mechanism Learning, as well as the availability of readily usable frameworks for differentially private learning make it feasible to apply the following result to mechanism learning: Differential privacy “can ensure that participants have limited effect on the outcome of the mechanism, and as a consequence have limited incentive to lie.”
The code for our ICML DPML Workshop paper implements a prototype for approximately truthful multi-item multi-bidder auctions without a requirement to know the valuation distributions of the agents a priori.
The current iteration is a JAX implementation based on a GAN-like architecture. Integration of DP will follow soon.
When differential privacy in the local model is used as a data privacy technique in a distributed computation, participants guard their own privacy when executing a DP mechanism correctly, thus they have a strong motivation to do so.
When differential privacy is used in mechanism design, it usually bounds the influence of each participant on the outcome, thus they have a motivation to not execute it correctly. Using a trusted curator (central model) leads to the issues that come with having a single trusted node.
A solution to this issue can be to distribute the execution of the central model of differential privacy. Using secure multiparty computation, the curator gets distributed over multiple computational entities, of which some can be compromised, while still keeping certain integrity guarantees. This way, neither the clients need to be trusted, nor does the mechanism rely on a single trusted party.
I have integrated the secure summation of libprio with the discrete gaussian of googles differential privacy library to provide this. Most of the research problems have been solved, this project aims to achieve a robust, scalable and secure implementation. A port based on OpenDP and libprio-rs and and integration into JAX are the next steps.
TCN is an open protocol for decentralized, privacy preserving contact tracing. It is developed by the TCN Coalition. I contributed to the analysis of its threatmodel and scaling behaviour.