Privacy Protection Amplified: Leveraging Agent-Based Simulation

Abstract

Privacy protection is inherently complex and challenging, especially with the rise of interdependent data sharing. This dissertation addresses the intricate nature of privacy risks by developing agent-based models to simulate user behavior in sensitive contexts, focusing on genomic data sharing and health data auditing. In the first part, we model the social dynamics of genomic data sharing using a game-theoretic framework. Our analysis highlights how individual decisions can produce network effects and negative externalities, leading to multiple Nash equilibria and potential tragedies of the commons. In the second part, we design a reinforcement learning–based auditing policy for online data sharing and access platforms, such as the NIH’s All of Us program. By integrating agent-based modeling with deep reinforcement learning over a dynamic bipartite graph of users and workspaces, our approach learns effective auditing strategies that leverage peer effects. Together, these contributions advance our understanding of interdependent privacy and provide a new simulation-based paradigm for designing privacy-preserving systems.