Minutemen is a group of graduate students at UMass who conduct research on differential privacy through their research, they found the temporal and geographical aspect of the 2020 Differential Privacy Temporal Map Challenge very interesting. The team members are Ryan McKenna, Joie Wu, Arisa Tajima, Brett Mullins, Siddhant Pradhan, and Cecilia Ferrando.
The team developed an adaptive-grid style mechanism to effectively deal with the imbalanced data. They began by invoking the analytic Gaussian mechanism to measure all 1-way marginals. Then, they inspected the noisy histograms and determined which parts of the domain could be measured at a finer granularity. Then, they invoked the analytic Gaussian mechanism again to measure cells in the noisy 2-way marginals that could have a sufficiently large count, which is decided based on the noisy 1-way marginals. The team repeated this for 3- and 4-way marginals. After measuring all of these statistics, they post-process the measurements using Private-PGM, which constructs a graphical model that preserves the measured information well, and they generated synthetic data from that model. The chart below depicts this process:
After measuring the 1-way marginals for column 1 and column 2, we inspect the noisy counts and identify cells whose count is above a given threshold (in this case, 100). If the count for particular values in column 1 and column 2 are both above the threshold, then we will measure the number of records that have both of those values at the same time. In the figure above, the cells in the 2-way marginal that we measure are color-coded in green.
To contact this team, please email Ryan McKenna at rmckenna [at] cs.umass.edu (rmckenna[at]cs[dot]umass[dot]edu).