Instantiate the consortium-level neighboring-world model

Determine an appropriate class of data-generating distributions and scheduler behaviors, denoted by \(\Theta\), for applying the institution-level Pufferfish participation-privacy definition to a scientific consortium.

Background

The paper formulates institution-level participation privacy using a Pufferfish-style definition. The neighboring secrets concern whether a particular institution contributed during a specified set of rounds, while the observer’s view includes released model weights, timing, and orchestration metadata. Unlike record-level differential privacy, the model must capture correlations between an institution’s data and its scheduling behavior. The authors state the target definition but do not provide a suitable scientific-consortium instantiation of the distribution class $\Theta$.