Mean-Field Game Approach to Privacy in Federated Learning
A recent study published on arXiv (2607.23029) introduces a mean-field privacy game aimed at enhancing privacy in federated learning. Within this model, each client makes strategic decisions regarding their privacy budget while engaging with the broader population via a single mean-field statistic. This method results in a manageable Nash equilibrium applicable to any number of clients, supports diverse client preferences, and provides an exponentially diminishing privacy assurance based on a log-Sobolev inequality. This research connects previous studies on noise injection and multi-agent games, presenting a scalable approach for secure collaborative model training.
Key facts
- Paper arXiv:2607.23029 proposes mean-field privacy game for federated learning
- Each client strategically chooses its own privacy budget
- Interaction occurs through a single mean-field statistic
- Yields tractable Nash equilibrium for arbitrarily many clients
- Accommodates heterogeneous client preferences
- Inherits exponentially decaying privacy guarantee via log-Sobolev inequality
- Bridges noise injection and multi-agent game approaches
- Published on arXiv
Entities
Institutions
- arXiv