A Non-Disclosure Incentive Mechanism for Federated Learning Via Alternating-offers Bargaining

In this paper, we propose a non-disclosure incentive mechanism for federated learning via alternating-offers bargaining which does not utilize device information of clients while improving learning performance and approximating an optimal resource allocation. By defining utility functions of clients that reflect their self estimated performances and determining minimum compensation information, clients can be selected for participation without revealing their device information such as loss of their local model or dataset size. To determine the compensation resource allocation, server and clients perform alternating-offers bargaining, which does not require sharing their utility functions. Experiment results show that the proposed incentive mechanism can speed up convergence, improve test accuracy, and induce compensation resource allocation near the Nash bargaining soution while not revealing the device information of clients.