Federated learning based on dynamic hierarchical game incentives in Industrial Internet of Things

Federated Learning (FL) can effectively protect user data privacy while performing distributed machine learning, which has shown a mighty capability to safely train intelligent industrial models on the Industrial Internet of Things (IIoT). However, in real scenarios, the performance of the global model is threatened by the possibility of IIoT devices launching malicious attacks; at the same time, it is difficult for devices to actively participate in the FL process without sufficient utility, resulting in a model that is not sufficiently data-driven. In this paper, we propose a dynamic hierarchical game incentive mechanism to achieve secure and fair FL in IIoT. Specifically, we design a cloud-factory-device three-layer collaborative FL architecture. In the cloud-factory layer, we design each iteration of FL as a cooperative game process and propose a reward allocation scheme based on the Shapley value to accomplish the incentive process. After theoretical deduction, we demonstrate that the scheme is consistent with rationality, fairness, and additionality. In the factory-device layer, we construct the problem for the FL iteration process and model this problem as a two-stage Stackelberg game process. We design a reputation-based adaptive local FL iteration algorithm and a non-cooperative game- based incentive mechanism to solve the game process. While solving the problem, we further consider the device's dynamics and delay time to fit the real IIoT scenario. Theoretical analyses and experimental results show that our proposed mechanism has good performance.