Lecture 2: Privacy as a Desideratum in Mechanism Design -- Making the Exponential Mechanism Truthful

For those of you following the privacy and mechanism design course at home, tomorrow's Lecture 2 is now online.

This lecture explores a different aspect of privacy and mechanism design. Last week, we considered the use of differential privacy purely as a tool, to design a prior free asymptotically truthful and revenue optimal mechanism for digital goods auctions. In this lecture, we motivate differential privacy as a desideratum for its own sake, and argue that it gives mechanisms (often at very little cost) a natural notion of robustness against the possibility of extra unmodeled parts of agents utility functions.

Finally, we derive a private version of the VCG mechanism, that for general social choice problems is simultaneously:

1. (Exactly) dominant strategy truthful
2. Differentially private
3. (Approximately) welfare optimal

This result is from Huang and Kannan 2012: "The Exponential Mechanism for Social Welfare: Private, Truthful, and Nearly Optimal". Although this is a recent paper, it is a basic, important, foundational result that "should" have been discovered earlier. It fits in very nicely at the beginning of the semester.

(And for those of you actually taking this class and looking for inspiration to do a course project, this paper started as Zhiyi Huang's course project in the 2011 privacy course I taught here at Penn)