Space–air–ground integrated network (SAGIN) has shown strong communication and computation abilities in various Internet of Things (IoTs) applications with the assistance of artificial intelligence (AI), such as emergency communication and remote sensing. However, resource heterogeneity of aerial devices is always the bottleneck of the performance of AI models and energy efficiency. In this paper, a collaborative federated learning (FL) scheme based on device-to-device (D2D) communication in unmanned aerial vehicle (UAV)-assisted SAGIN is proposed to address the issue of heterogeneity. Aerial devices with limited communication and computation resource can offload partial nonprivacy data samples to proximity D2D pair, which can assist to train FL models. An optimization problem is proposed to minimize the total energy consumption and the loss function of local FL models. In order to solve the mixed integer nonlinear problem (MINLP), a data offloading selection strategy based on proximity discovery and an iterative method-based resource allocation algorithm (IRA) are proposed. In addition, the closed-form solutions of the optimized variables are obtained. Simulation results demonstrate that the proposed collaborative training scheme based on D2D can reduce the impact of heterogeneity on FL model performance and IRA can effectively reduce energy consumption while simultaneously enhancing training efficiency of FL.

The space–air–ground information network (SAGIN) has been widely used due to its excellent performances including wide coverage and high flexibility. However, the dynamic network topology of SAGIN presents challenges for traditional protocols. The statistical priority-based multiple access (SPMA) control protocol has received widespread attention because it effectively allocates resources in networks with heterogeneous terminals and dynamic topology. However, the existing SPMA protocols suffer from issues like large errors and low prediction accuracy in channel load statistics. Therefore, this paper proposes an improved SPMA based on the bi-directional long short-term memory (BiLSTM) neural network. First, we analyze and correct errors in channel load statistics at the physical layer, then develop a BiLSTM-based channel load prediction model, and finally simulated the improved SPMA using Matlab. Experimental results show that the proposed channel load prediction model achieves good prediction accuracy, and the improved SPMA protocol markedly improves channel utilization, providing differentiated services for multi-priority businesses.