Privacy-Preserving Incentive Mechanism Design for Wireless Federated Learning

Federated learning (FL) is an emerging technique with privacy preserving capabilities to train machine learning models in a distributed environment. However, the selfish nature of edge devices (EDs) may limit the performance of FL, which are unwilling to contribute their data and power to help model training without any reward. Designing an efficient incentive mechanism is necessary to motivate the devices to participate in FL. This paper proposes a new privacy-preserving incentive power allocation mechanism for wireless FL, while considering model uploading error and the vulnerability of model privacy to eavesdropping attacks. The interaction between the edge server and EDs is modeled as a two-stage Stackelberg game to improve both of their revenues, which are affected by power consumption, model accuracy, model security and training latency. By analyzing the game equilibrium, the transmit powers of model uploading and cooperative jamming of EDs, as well as the reward strategy of the server are iteratively optimized. Simulation results show that the proposed method achieves higher utilities than the benchmarks, validating its effectiveness in improving the power utilization of EDs.