Reproduce “Privacy as Commodity: MFG-RegretNet for Large-Scale Privacy Trading in Federated Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kangkang Sun, Jianhua Li, Xiuzhen Chen, Weizhi Meng, Minyi Guo

MFG-RegretNet introduces a scalable deep learning-based auction mechanism for federated learning, enabling privacy budget trading where clients are compensated for their personalized differential privacy contributions. The framework demonstrates superior approximate incentive compatibility and auction efficiency with near-linear scalability for large client populations, all while maintaining competitive FL model accuracy.

Key insights

  • Personalized differential privacy budgets can be formalized as a tradable commodity within an auction framework, creating an economic incentive for client participation in Federated Learning.
  • Mean-Field Games provide a theoretically sound method to approximate equilibrium in large-scale N-player auction games, offering significant scalability benefits for distributed systems like Federated Learning.
  • Differentiable mechanism design, specifically MFG-RegretNet, can learn complex, incentive-compatible auction rules from data, effectively adapting to unknown and heterogeneous client valuation distributions while leveraging MFG for scalability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.