The imminent deployment of sixth-generation (6G) wireless communication systems promises new opportunities and challenges for model training using data from the edge devices in the Internet of Things (IoT). However, current research has yet to fully address the efficiency and scalability challenges arising from the extensive connectivity of edge devices across various scenarios. The presence of malicious devices further intensifies the system uncertainty during large-scale data interactions and model training, making it difficult for a single model to effectively manage the complexities introduced by heterogeneous devices and dynamic network conditions. To overcome these challenges, we propose FedSC, an innovative edge computing framework that leverages side-chain technology for efficient edge node management and employs federated learning to enable robust cross-device and cross-scenario model interactions. To accelerate the multimodel aggregation process, we introduce an asynchronous cross-domain iterative algorithm (ACDI) based on smart contracts. Additionally, to mitigate the impact of malicious and inactive nodes, we propose a robust consensus algorithm and a committee mechanism for leader node election based on contribution value. Experimental results demonstrate that the proposed FedSC achieves a 3.2% and 44.23% accuracy improvement on i.i.d. and non-i.i.d. dataset, respectively, along with a remarkable latency reduction of 256.51%, compared to FedAvg. Our work is conducive to the training of multiple models in different IoT scenarios, utilizing substantial amounts of IoT device data and facilitating collaboration between models. Furthermore, it enables the provision of fundamental services to diverse applications in 6G.

Federated learning (FL) is a distributed privacy-preserving paradigm of machine learning that enables efficient and secure model training through the collaboration of multiple clients. However, imperfect channel estimation and resource constraints of edge devices severely hinder the convergence of typical wireless FL, while the trade-off between communications and computation still lacks in-depth exploration. These factors lead to inefficient communications and hinder the full potential of FL from being unleashed. In this regard, we formulate a joint optimization problem of communications and learning in wireless networks subject to dynamic channel variations. For addressing the formulated problem, we propose an integrated adaptive$n$-ary compression and resource management framework (ANC) that is capable of adjusting the selection of edge devices and compression schemes, and allocates the optimal resource blocks and transmit power to each participating device, which effectively improves the energy efficiency and scalability of FL in resource-constrained environments. Furthermore, an upper bound on the expected global convergence rate is derived in this paper to quantify the impacts of transmitted data volume and wireless propagation on the convergence of FL. Simulation results demonstrate that the proposed adaptive framework achieves much faster convergence while maintaining considerably low communication overhead.