ABSTRACT Deploying Federated Learning (FL) on IoT edge devices faces three major challenges: computation, communication, and non‐independent and identically distributed (Non‐IID) data. We observe that the core attention mechanisms in existing lightweight networks have serious flaws in the federated setting—specifically, parameter redundancy and a lack of spatial modeling capabilities. These issues directly increase the communication overhead of federated learning and hinder its adaptation to heterogeneous data, becoming a key technical bottleneck. To overcome this bottleneck, we have designed a novel lightweight network architecture called FEMBNet. The core of this architecture is the FEMA module, specifically optimized for the federated setting. It can replace inefficient traditional attention mechanisms in the federated learning scenario. Experimental results show that in Non‐IID scenarios, FEMBNet can improve global model accuracy by 3.28% while reducing 13.87 k redundant parameters in its core mechanism. This work provides a proven architectural solution for efficient FL in IoT systems, demonstrating that network structure optimization can effectively address data heterogeneity and communication bottlenecks.
Tian et al. (Fri,) studied this question.