Incentive Mechanism for Federated Learning Participants Based on Statistical Analysis Features
Hao Li, Kai Shi · 2023
Compared to traditional machine learning approaches, Google's innovative framework for federated learning has emerged as a prominent focus within the realm of collaborative efforts involving multi-source data and multiple users, all aimed at enhancing model performance. Distinguished by its high learning efficiency and robust privacy-preserving capabilities, this framework has garnered significant attention. In the context of cross-device federated learning, potential challenges can stem from participants who display a lack of enthusiasm for engaging in model training or who meticulously manipulate their submitted model parameters. These malicious behaviors carry the potential to undermine the integrity of the global model, thereby exerting a notable influence on the trajectory of the global model training process.To tackle these challenges, this paper introduces an incentive mechanism tailored for federated learning participants, grounded in the analysis of statistical features. By harnessing insights from statistical features, this mechanism dynamically adjusts participants' weights, thereby curbing the impact of malicious users. Under the assumption of rational participant behavior, the mechanism not only mitigates potential harm but also encourages proactive involvement. Empirical findings underscore the effectiveness of this adaptive mechanism in constraining the behavior of malicious participants in federated learning, thereby fortifying protective measures for the global model. The practical utility of this mechanism is demonstrated convincingly.