How to Budget Privacy in Federated Learning? A Correlated Equilibrium Perspective

Facilitated by mobile edge computing (MEC), federated learning (FL) can be deployed at the network edge to protect the local data privacy of mobile devices for real-time data-driven applications. To further prevent privacy inference from local models, FL has been enhanced by differential privacy (DP) technology. The existing DP-enhanced FL studies focus on developing DP mechanisms and designing incentive schemes for DP implementation in FL, overlooking the issue of fair privacy budgeting among all clients. To fill this gap, we achieve fairness by balancing the privacy budget, model performance, and cost via defining a privacy budget determination (PBD) game among clients with similar local models. However, it is nontrivial to achieve this balance due to two challenges: the interdependency among clients regarding choosing privacy budgets and the unavailability of other devices' parameters. To resolve the first challenge, we employ the concept of correlated equilibrium (CE) to capture the mutual influences among clients in deciding privacy budgets. We then formulate an optimization problem with CE and individual rationality (IR) constraints to derive the optimal decision. For the second challenge, we exploit the centralized position of the aggregation server to help calculate the solution to the optimization problem. Experiments on real-world datasets demonstrate that the proposed mechanism ensures fairness while achieving superior client utility and global model performance.