Authors: Jan Van Bruggen, Cutter Coryell, James Chang
Model for the privacy of streamed content with respect to complexity, interactivity, and noisiness
We have structured our model as the canonical privacy problem, with a curator that protects a database from an adversary capable of making repeated query attacks in attempt to reconstruct the database.
The only database model we are analyzing is an interactive content network of nodes connected by links with various utilities. At each node, user input (or a user response to a question) can change the link utilities, so link utility depends on both the nodes and the user input.
The curator processes all analyst queries to the database and adds noise by calculating link probabilities according to the exponential mechanism, instead of simply normalizing the link utilities to obtain link probabilities.
The adversary reconstructs the database by analyzing the sequence obtained by querying through the curator. We implement a simple reconstruction algorithm where the adversary uses multiplicative weights to increase the utility of each link seen in the sequence. We evaluate the adversary's success with the KL-Divergence of the probability distributions (over sequences) that represent the curator's database and the adversary's database.