Cloud–edge–end computing architecture is crucial for large-scale edge data processing and analysis. However, the diversity of terminal nodes and task complexity in this architecture often result in non-independent and identically distributed (non-IID) data, making it challenging to balance data heterogeneity and privacy protection. To address this, we propose a privacy-preserving federated learning method based on cloud–edge–end collaboration. Our method fully considers the three-tier architecture of cloud–edge–end systems and the non-IID nature of terminal node data. It enhances model accuracy while protecting the privacy of terminal node data. The proposed method groups terminal nodes based on the similarity of their data distributions and constructs edge subnetworks for training in collaboration with edge nodes, thereby mitigating the negative impact of non-IID data. Furthermore, we enhance WGAN-GP with attention mechanism to generate balanced synthetic data while preserving key patterns from original datasets, reducing the adverse effects of non-IID data on global model accuracy while preserving data privacy. In addition, we introduce data resampling and loss function weighting strategies to mitigate model bias caused by imbalanced data distribution. Experimental results on real-world datasets demonstrate that our proposed method significantly outperforms existing approaches in terms of model accuracy, F1-score, and other metrics.

The Space-Air-Ground Integrated Network (SAGIN) has emerged as a core architecture for future intelligent communication due to its wide-area coverage and dynamic heterogeneous characteristics. However, its high latency, dynamic topology, and privacy–security challenges severely constrain the application of Federated Learning (FL). This paper proposes a Privacy-Preserving Federated Learning framework for SAGIN (PPFL-SAGIN), which for the first time integrates differential privacy, adaptive transfer learning, and bi-level reinforcement learning to systematically address data heterogeneity, device dynamics, and privacy leakage in SAGINs. Specifically, (1) an adaptive knowledge-sharing mechanism based on transfer learning is designed to balance device heterogeneity and data distribution divergence through dynamic weighting factors; (2) a bi-level reinforcement learning device selection strategy is proposed, combining meta-learning and hierarchical attention mechanisms to optimize global–local decision-making and enhance model convergence efficiency; (3) dynamic privacy budget allocation and robust aggregation algorithms are introduced to reduce communication overhead while ensuring privacy. Finally, experimental evaluations validate the proposed method. Results demonstrate that PPFL-SAGIN significantly outperforms baseline solutions such as FedAvg, FedAsync, and FedAsyncISL in terms of model accuracy, convergence speed, and privacy protection strength, verifying its effectiveness in addressing privacy preservation, device selection, and global aggregation in SAGINs.

With the proliferation of LEO satellite constellations and increasing demands for on-orbit intelligence, satellite networks generate massive, heterogeneous, and privacy-sensitive data. Ensuring efficient model collaboration under strict privacy constraints remains a critical challenge. This paper proposes HiSatFL, a cross-domain adaptive and privacy-preserving federated learning framework tailored to the highly dynamic and resource-constrained nature of satellite communication systems. The framework incorporates an orbital-aware hierarchical FL architecture, a multi-level domain adaptation mechanism, and an orbit-enhanced meta-learning strategy to enable rapid adaptation with limited samples. In parallel, privacy is preserved via noise-calibrated feature alignment, differentially private adversarial training, and selective knowledge distillation, guided by a domain-aware dynamic privacy budget allocation scheme. We further establish a unified optimization framework balancing privacy, utility, and adaptability, and derive convergence bounds under dynamic topologies. Experimental results on diverse remote sensing datasets demonstrate that HiSatFL significantly outperforms existing methods in accuracy, adaptability, and communication efficiency, highlighting its practical potential for collaborative on-orbit AI.

With the development of 6G communication technology, Space–Air–Ground Integrated Networks (SAGINs) have become critical infrastructure for global intelligent collaborative computing. However, federated learning deployment in SAGINs faces three severe challenges: the high dynamics of satellite orbital motion, node resource heterogeneity, and privacy vulnerabilities in data transmission. This paper proposes an adaptive compressed sensing differential privacy federated learning framework based on orbital spatiotemporal characteristics. First, we design orbital periodicity-driven time-varying sparse sensing matrices that dynamically adjust compression strategies according to satellite orbital positions, achieving intelligent communication efficiency optimization. Second, we propose an orbital predictability-based privacy budget temporal allocation mechanism and perform differential privacy noise injection in the compressed domain, establishing a compression–privacy joint optimization algorithm. Furthermore, we construct an energy–communication–privacy ternary collaborative mechanism that achieves multi-objective dynamic balance through model predictive control. Finally, we design reinforcement learning-based dynamic routing scheduling and hierarchical aggregation strategies to effectively handle the time-varying characteristics of network topology. Simulation experiments demonstrate that compared to existing methods, the proposed approach achieves 3–12% improvement in model accuracy and 30–50% enhancement in communication efficiency while maintaining differential privacy protection with dynamic privacy budget $\epsilon \in [0.1,10.0]$ and compression ratio $\rho \in [0.2,0.8]$. Unlike static compressed sensing approaches that ignore orbital periodicity, the proposed orbital-driven time-varying sensing matrices reduce reconstruction error by up to 19.4% compared to fixed-matrix baselines, validating the synergistic effectiveness of integrating orbital spatiotemporal characteristics with federated learning in 6G SAGIN deployments. The framework assumes reliable orbital propagation via SGP4/SDP4 models and does not account for Doppler frequency shifts or inter-satellite link handover delays; future extensions include scalability to mega-constellations and integration of quantum-resistant privacy mechanisms.