Dual-layer grouping consensus for decentralized federated learning

Abstract Federated learning has a reliance on centralized coordination, which brings several limitations (e.g. performance bottlenecks and single-point failures) and motivates the development of decentralized federated learning (DFL). In DFL, consensus mechanism plays a key role in ensuring reliable decentralized operation. Yet, existing consensus methods still struggle to balance multiple performance metrics and also face challenges such as insufficient diversity in consensus models. To address these challenges, the study proposes a dual-layer grouping consensus mechanism to balance various performance metrics, including accuracy and efficiency, while improving model diversity. Specifically, the implementation of the dual-layer grouping consensus mechanism is facilitated by a self-organizing grouping method to optimize group formation and a differentiated incentive mechanism to evaluate client performance. The experiments conducted on three datasets demonstrate that the framework delivers distinct outcomes: the global model gains a 3.47% accuracy increase, whereas the Efficiency Group specifically cuts time overhead to 56.61% of the best-performing baseline.