Federated learning is a type of distributed machine learning in which models learn by using large-scale decentralized data between servers and devices. In a short-range wireless communication environment, it can be difficult to apply federated learning because the number of devices in one access point (AP) is small, which can be small enough to perform federated learning. Therefore, it means that the minimum number of devices required to perform federated learning cannot be matched by the devices included in one AP environment. To do this, we propose to obtain a uniform global model regardless of data distribution by considering the multi-AP coordination characteristics of IEEE 802.11be in a decentralized federated learning environment. The proposed method can solve the imbalance in data transmission due to the non-independent and identically distributed (non-IID) environment in a decentralized federated learning environment. In addition, we can also ensure the fairness of multi-APs and determine the update criteria for newly elected primary-APs by considering the learning training time of multi-APs and energy consumption of grouped devices performing federated learning. Thus, our proposed method can determine the primary-AP according to the number of devices participating in the federated learning in each AP during the initial federated learning to consider the communication efficiency. After the initial federated learning, fairness can be guaranteed by determining the primary-AP through the training time of each AP. As a result of performing decentralized federated learning using the MNIST and FMNIST dataset, the proposed method showed up to a 97.6% prediction accuracy. In other words, it can be seen that, even in a non-IID multi-AP environment, the update of the global model for federated learning is performed fairly.
Keywords: 802.11be; Wi-Fi 7; distributed machine learning; energy consumption optimization; federated learning; multi-AP coordination.