Understanding mean-field effects of large-population user data obfuscation in machine learning
Alex Dunyak, Quanyan Zhu · 2018
Recently, data-based services have bloomed, because of the prevalence of online tracking, wearable computing, and the Internet of Things. However, due to privacy concerns, users may use tools to obfuscate their data, rendering these services less useful. This conflict places the service's desire for accuracy, and the users' desire for both accuracy and privacy, in contention. We propose a game-theoretic model for this conflict. By promising bounds on how much data gets collected and sold, services can incentivize users to report their data truthfully. We model this incentivization process as a Stackelberg game with the service provider as the leader. The users react to privacy promises as Stackelberg followers and interact with each other by playing a mean-field game. By representing the users as an interval on the continuum, we reduce the computational complexity of finding Nash and Stackelberg equilibriums.